Some bug fixes and improvements for FK

There's still some glitches around selection and deletion.

Cargo.toml

@@ -23,6 +23,8 @@ members = [
 
   # Desktop apps
   "crates/desktop/src-tauri",
+  "crates/desktop/ml_models",
+  "crates/desktop/ml_weights_registry",
 
   # Shared libraries
   "crates/lib/actix_cors_configs",
@@ -30,6 +32,7 @@ members = [
   "crates/lib/api_clients/cloud_storage",
   "crates/lib/api_clients/google_drive_common",
   "crates/lib/api_clients/google_sign_in",
+  "crates/lib/api_clients/openai_sora_client",
   "crates/lib/api_clients/thumbnail_generator",
   "crates/lib/bootstrap",
   "crates/lib/bucket_paths",
@@ -56,6 +59,7 @@ members = [
   "crates/lib/files/mimetypes",
   "crates/lib/files/videos",
   "crates/lib/idempotency",
+  "crates/lib/image_algorithms",
   "crates/lib/jobs_common",
   "crates/lib/logging",
   "crates/lib/markdown",
@@ -115,3 +119,23 @@ members = [
   # see usage: https://crates.io/crates/cargo-hakari
   "crates/my-workspace-hack",
 ]
+
+[workspace.package]
+publish = false
+
+[workspace.dependencies]
+# Hakari workspace hack (faster compile times)
+my-workspace-hack = { path = "crates/my-workspace-hack" }
+
+# Internal libs
+errors = { path = "crates/lib/errors" }
+
+# External libs
+# Centrally pinned versions so that we can uniformly upgrade/downgrade versions across the workspace
+# If a crate needs to use an older/newer version, the dependency should be managed in that crate's own Cargo.toml
+anyhow = "1.0.97"
+log = "0.4.27"
+reqwest = { version = "0.12.15", features = ["json", "cookies"] }
+serde = "1.0.219"
+serde_derive = "1.0.219"
+serde_json = "1.0.140"

crates/_types/enums/src/by_table/generic_inference_jobs/inference_job_type.rs

@@ -40,6 +40,10 @@ pub enum InferenceJobType {
   #[serde(rename = "studio_gen2")]
   StudioGen2,
 
+  /// Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
+
   /// A job that turns "FBX" game engine files into "GLTF" files (Bevy-compatible).
   #[serde(rename = "convert_fbx_gltf")]
   ConvertFbxToGltf,
@@ -108,6 +112,7 @@ impl InferenceJobType {
       Self::GptSovits => "gpt_sovits",
       Self::ComfyUi => "comfy_ui",
       Self::StudioGen2 => "studio_gen2",
+      Self::ImageGenApi => "image_gen_api",
       Self::ConvertFbxToGltf => "convert_fbx_gltf",
       Self::MocapNet => "mocap_net",
       Self::RvcV2 => "rvc_v2",
@@ -132,6 +137,7 @@ impl InferenceJobType {
       "gpt_sovits" => Ok(Self::GptSovits),
       "comfy_ui" => Ok(Self::ComfyUi),
       "studio_gen2" => Ok(Self::StudioGen2),
+      "image_gen_api" => Ok(Self::ImageGenApi),
       "convert_fbx_gltf" => Ok(Self::ConvertFbxToGltf),
       "mocap_net" => Ok(Self::MocapNet),
       "rvc_v2" => Ok(Self::RvcV2),
@@ -159,6 +165,7 @@ impl InferenceJobType {
       Self::GptSovits,
       Self::ComfyUi,
       Self::StudioGen2,
+      Self::ImageGenApi,
       Self::ConvertFbxToGltf,
       Self::MocapNet,
       Self::RvcV2,
@@ -197,6 +204,7 @@ mod tests {
       assert_serialization(InferenceJobType::GptSovits, "gpt_sovits");
       assert_serialization(InferenceJobType::ComfyUi, "comfy_ui");
       assert_serialization(InferenceJobType::StudioGen2, "studio_gen2");
+      assert_serialization(InferenceJobType::ImageGenApi, "image_gen_api");
       assert_serialization(InferenceJobType::ConvertFbxToGltf, "convert_fbx_gltf");
       assert_serialization(InferenceJobType::MocapNet, "mocap_net");
       assert_serialization(InferenceJobType::RvcV2, "rvc_v2");
@@ -220,6 +228,7 @@ mod tests {
       assert_eq!(InferenceJobType::GptSovits.to_str(), "gpt_sovits");
       assert_eq!(InferenceJobType::ComfyUi.to_str(), "comfy_ui");
       assert_eq!(InferenceJobType::StudioGen2.to_str(), "studio_gen2");
+      assert_eq!(InferenceJobType::ImageGenApi.to_str(), "image_gen_api");
       assert_eq!(InferenceJobType::ConvertFbxToGltf.to_str(), "convert_fbx_gltf");
       assert_eq!(InferenceJobType::MocapNet.to_str(), "mocap_net");
       assert_eq!(InferenceJobType::RvcV2.to_str(), "rvc_v2");
@@ -243,6 +252,7 @@ mod tests {
       assert_eq!(InferenceJobType::from_str("gpt_sovits").unwrap(), InferenceJobType::GptSovits);
       assert_eq!(InferenceJobType::from_str("comfy_ui").unwrap(), InferenceJobType::ComfyUi);
       assert_eq!(InferenceJobType::from_str("studio_gen2").unwrap(), InferenceJobType::StudioGen2);
+      assert_eq!(InferenceJobType::from_str("image_gen_api").unwrap(), InferenceJobType::ImageGenApi);
       assert_eq!(InferenceJobType::from_str("convert_fbx_gltf").unwrap(), InferenceJobType::ConvertFbxToGltf);
       assert_eq!(InferenceJobType::from_str("mocap_net").unwrap(), InferenceJobType::MocapNet);
       assert_eq!(InferenceJobType::from_str("rvc_v2").unwrap(), InferenceJobType::RvcV2);
@@ -261,7 +271,7 @@ mod tests {
     fn all_variants() {
       // Static check
       let mut variants = InferenceJobType::all_variants();
-      assert_eq!(variants.len(), 16);
+      assert_eq!(variants.len(), 17);
       assert_eq!(variants.pop_first(), Some(InferenceJobType::VideoRender));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::LivePortrait));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::FaceFusion));
@@ -269,6 +279,7 @@ mod tests {
       assert_eq!(variants.pop_first(), Some(InferenceJobType::GptSovits));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::ComfyUi));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::StudioGen2));
+      assert_eq!(variants.pop_first(), Some(InferenceJobType::ImageGenApi));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::ConvertFbxToGltf));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::MocapNet));
       assert_eq!(variants.pop_first(), Some(InferenceJobType::RvcV2));

crates/_types/enums/src/by_table/generic_inference_jobs/inference_model_type.rs

@@ -33,6 +33,10 @@ pub enum InferenceModelType {
   #[serde(rename = "seed_vc")]
   SeedVc,
 
+  /// NB: This is for Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
+
   #[serde(rename = "tacotron2")]
   Tacotron2,
   #[serde(rename = "vits")]
@@ -70,6 +74,7 @@ impl InferenceModelType {
       Self::VallEX => "vall_e_x",
       Self::RerenderAVideo => "rerender_a_video",
       Self::StableDiffusion => "stable_diffusion",
+      Self::ImageGenApi => "image_gen_api",
       Self::SeedVc => "seed_vc",
       Self::MocapNet => "mocap_net",
       Self::StyleTTS2 => "styletts2",
@@ -90,6 +95,7 @@ impl InferenceModelType {
       "vall_e_x" => Ok(Self::VallEX),
       "rerender_a_video" => Ok(Self::RerenderAVideo),
       "stable_diffusion" => Ok(Self::StableDiffusion),
+      "image_gen_api" => Ok(Self::ImageGenApi),
       "mocap_net" => Ok(Self::MocapNet),
       "styletts2" => Ok(Self::StyleTTS2),
       "comfy_ui" => Ok(Self::ComfyUi),
@@ -113,6 +119,7 @@ impl InferenceModelType {
       Self::VallEX,
       Self::RerenderAVideo,
       Self::StableDiffusion,
+      Self::ImageGenApi,
       Self::MocapNet,
       Self::StyleTTS2,
       Self::ConvertFbxToGltf,
@@ -140,6 +147,7 @@ mod tests {
       assert_serialization(InferenceModelType::VallEX, "vall_e_x");
       assert_serialization(InferenceModelType::RerenderAVideo, "rerender_a_video");
       assert_serialization(InferenceModelType::StableDiffusion, "stable_diffusion");
+      assert_serialization(InferenceModelType::ImageGenApi, "image_gen_api");
       assert_serialization(InferenceModelType::MocapNet, "mocap_net");
       assert_serialization(InferenceModelType::ComfyUi, "comfy_ui");
       assert_serialization(InferenceModelType::StyleTTS2, "styletts2");
@@ -159,6 +167,7 @@ mod tests {
       assert_eq!(InferenceModelType::VallEX.to_str(), "vall_e_x");
       assert_eq!(InferenceModelType::RerenderAVideo.to_str(), "rerender_a_video");
       assert_eq!(InferenceModelType::StableDiffusion.to_str(), "stable_diffusion");
+      assert_eq!(InferenceModelType::ImageGenApi.to_str(), "image_gen_api");
       assert_eq!(InferenceModelType::MocapNet.to_str(), "mocap_net");
       assert_eq!(InferenceModelType::StyleTTS2.to_str(), "styletts2");
       assert_eq!(InferenceModelType::ConvertFbxToGltf.to_str(), "convert_fbx_gltf");
@@ -176,6 +185,7 @@ mod tests {
       assert_eq!(InferenceModelType::from_str("vall_e_x").unwrap(), InferenceModelType::VallEX);
       assert_eq!(InferenceModelType::from_str("rerender_a_video").unwrap(), InferenceModelType::RerenderAVideo);
       assert_eq!(InferenceModelType::from_str("stable_diffusion").unwrap(), InferenceModelType::StableDiffusion);
+      assert_eq!(InferenceModelType::from_str("image_gen_api").unwrap(), InferenceModelType::ImageGenApi);
       assert_eq!(InferenceModelType::from_str("mocap_net").unwrap(), InferenceModelType::MocapNet);
       assert_eq!(InferenceModelType::from_str("styletts2").unwrap(), InferenceModelType::StyleTTS2);
       assert_eq!(InferenceModelType::from_str("comfy_ui").unwrap(), InferenceModelType::ComfyUi);
@@ -187,17 +197,18 @@ mod tests {
     fn all_variants() {
       // Static check
       let mut variants = InferenceModelType::all_variants();
-      assert_eq!(variants.len(), 13);
+      assert_eq!(variants.len(), 15);
       assert_eq!(variants.pop_first(), Some(InferenceModelType::ComfyUi));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::RvcV2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::SadTalker));
-      assert_eq!(variants.pop_first(), Some(InferenceModelType::SeedVc));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::SoVitsSvc));
+      assert_eq!(variants.pop_first(), Some(InferenceModelType::SeedVc));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::Tacotron2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::Vits));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::VallEX));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::RerenderAVideo));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::StableDiffusion));
+      assert_eq!(variants.pop_first(), Some(InferenceModelType::ImageGenApi));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::MocapNet));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::StyleTTS2));
       assert_eq!(variants.pop_first(), Some(InferenceModelType::ConvertFbxToGltf));

crates/_types/enums_public/src/by_table/generic_inference_jobs/public_inference_job_type.rs

@@ -91,6 +91,10 @@ pub enum PublicInferenceJobType {
   /// Re-render a video is a video style transfer algorithm. We developed code
   /// around it, but chose to develop AnimateDiff / ComfyUI support instead.
   RerenderAVideo,
+
+  /// Sora GPT 4o image gen
+  #[serde(rename = "image_gen_api")]
+  ImageGenApi,
 }
 
 /// NB: Legacy API for older code.
@@ -119,6 +123,7 @@ impl PublicInferenceJobType {
       InferenceJobType::Unknown => Self::Unknown,
       InferenceJobType::BevyToWorkflow => Self::BevyToWorkflow,
       InferenceJobType::RerenderAVideo => Self::RerenderAVideo,
+      InferenceJobType::ImageGenApi => Self::ImageGenApi,
     }
   }
   
@@ -140,6 +145,7 @@ impl PublicInferenceJobType {
       Self::SeedVc => InferenceJobType::SeedVc,
       Self::SoVitsSvc => InferenceJobType::SoVitsSvc,
       Self::StableDiffusion => InferenceJobType::StableDiffusion,
+      Self::ImageGenApi => InferenceJobType::ImageGenApi,
       Self::StyleTTS2 => InferenceJobType::StyleTTS2,
       Self::Tacotron2 => InferenceJobType::Tacotron2,
       Self::Unknown => InferenceJobType::Unknown,

crates/desktop/ml_models/Cargo.toml

@@ -0,0 +1,80 @@
+[package]
+name = "ml_models"
+edition = "2021"
+version = "0.0.1"
+authors = ["storyteller"]
+publish = false
+description = ""
+license = ""
+repository = ""
+
+[lib]
+name = "ml_models"
+path = "src/lib.rs"
+crate-type = ["staticlib", "cdylib", "rlib"]
+
+[package.metadata.cargo-udeps.ignore]
+normal = ["my-workspace-hack"]
+
+[dependencies]
+
+# Internal
+filesys = { path = "../../lib/files/filesys" }
+image_algorithms = { path = "../../lib/image_algorithms" }
+memory_caching = { path = "../../lib/caching/memory_caching" }
+ml_weights_registry = { path = "../ml_weights_registry" }
+primitives = { path = "../../lib/data/primitives" }
+
+# Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
+cudarc = "0.13.7"
+
+# External
+anyhow = "1.0"
+base64 = "0.22.1"
+bytes = "1.10"
+candle-core = { path = "../../vendor/candle/candle-core", features = ["cuda"] }
+candle-nn = { path = "../../vendor/candle/candle-nn", features = ["cuda"] }
+candle-transformers = { path = "../../vendor/candle/candle-transformers", features = ["cuda"] }
+const_format = "0.2.34"
+directories = "6.0.0"
+dirs = "6.0.0"
+futures = "0.3"
+half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
+hf-hub = "0.3.2"
+image = "0.25.5"
+log = "0.4"
+ndarray = "0.15.6"
+nvidia-checker = "0.1.2"
+once_cell = "1.20.3"
+ort = { version = "1.16.2", features = ["cuda"] }
+rand = "0.8.5"
+# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
+reqwest = { version = "0.12.12", features = ["stream"], default-features = false }
+serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"
+serde_yaml = "0.9"
+tauri = { version = "2.2.4", features = [] }
+tauri-plugin-log = "2"
+tempdir = "0.3.7"
+tempfile = "3.18.0"
+tokenizers = "0.15.2"
+tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
+tokio-util = { version = "0.7", features = ["codec"] }
+
+# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
+
+[features]
+default = ["cuda"]
+cuda = [
+    "candle-core/cuda",
+    "candle-nn/cuda",
+    "candle-transformers/cuda",
+    # NB(bt,2025-03-10): Building without Flash attention on Windows for now to bypass linker issues
+    "candle-transformers/flash-attn",
+    "candle-core/cudnn"
+]
+
+[dev-dependencies]
+libloading = "0.8.6"
+serial_test = "3.2.0"
+assertor = "0.0.3"

crates/desktop/ml_models/src/lib.rs

@@ -0,0 +1 @@
+pub mod ml;

crates/desktop/ml_models/src/ml/background_removal/mod.rs

@@ -0,0 +1,2 @@
+pub mod onnx_session;
+pub mod remove_image_background;

crates/desktop/ml_models/src/ml/background_removal/onnx_session.rs

@@ -1,8 +1,9 @@
 //! Adapted from https://github.com/dnanhkhoa/rust-background-removal/blob/main/src/onnx.rs
 
+use std::path::Path;
 use ort::{Environment, ExecutionProvider, GraphOptimizationLevel, SessionBuilder};
 
-pub fn onnx_session(onnx_model_file: &str) -> Result<ort::Session, anyhow::Error> {
+pub fn onnx_session(onnx_model_file: &Path) -> Result<ort::Session, anyhow::Error> {
   let environment = Environment::default().into_arc();
   let session = SessionBuilder::new(&environment)?
     .with_optimization_level(GraphOptimizationLevel::Level1)? // Configure model optimization level

crates/desktop/ml_models/src/ml/background_removal/remove_image_background.rs

@@ -0,0 +1,66 @@
+use crate::ml::background_removal::onnx_session::onnx_session;
+use image::{imageops, DynamicImage};
+use ndarray::{Array, CowArray};
+use ort::Value;
+use std::path::Path;
+use image_algorithms::remove_image_border::remove_image_border;
+
+pub async fn remove_image_background<P: AsRef<Path>>(model_path: P, image: DynamicImage) -> anyhow::Result<DynamicImage> {
+  // NB: Real time needed cuDNN 9, but this uses cuDNN 8:
+  // sudo apt-get install libcudnn8-dev libcudnn8
+  // update-alternatives: warning: forcing reinstallation of alternative /usr/include/x86_64-linux-gnu/cudnn_v9.h because link group libcudnn is broken
+  // update-alternatives: using /usr/include/x86_64-linux-gnu/cudnn_v8.h to provide /usr/include/cudnn.h (libcudnn) in manual mode
+  let session = onnx_session(model_path.as_ref())?;
+  let image = process_dynamic_image(&session, image)?;
+  
+  // Crop transparent border cruft away.
+  let image = remove_image_border(&image)?;
+  
+  Ok(image)
+}
+
+fn process_dynamic_image(
+  session: &ort::Session,
+  dynamic_img: DynamicImage,
+) -> Result<DynamicImage, anyhow::Error> {
+  let input_shape = session.inputs[0]
+    .dimensions()
+    .map(|dim| dim.unwrap())
+    .collect::<Vec<usize>>();
+  let input_img = dynamic_img.into_rgba8();
+  let scaling_factor = f32::min(
+    1., // Avoid upscaling
+    f32::min(
+      input_shape[3] as f32 / input_img.width() as f32, // Width ratio
+      input_shape[2] as f32 / input_img.height() as f32, // Height ratio
+    ),
+  );
+  let mut resized_img = imageops::resize(
+    &input_img,
+    input_shape[3] as u32,
+    input_shape[2] as u32,
+    imageops::FilterType::Triangle,
+  );
+  let input_tensor = CowArray::from(
+    Array::from_shape_fn(input_shape, |indices| {
+      let mean = 128.;
+      let std = 256.;
+
+      (resized_img[(indices[3] as u32, indices[2] as u32)][indices[1]] as f32 - mean) / std
+    })
+      .into_dyn(),
+  );
+  let inputs = vec![Value::from_array(session.allocator(), &input_tensor)?];
+  let outputs = session.run(inputs)?;
+  let output_tensor = outputs[0].try_extract::<f32>()?;
+  for (indices, alpha) in output_tensor.view().indexed_iter() {
+    resized_img[(indices[3] as u32, indices[2] as u32)][3] = (alpha * 255.) as u8;
+  }
+  let output_img = imageops::resize(
+    &resized_img,
+    (input_img.width() as f32 * scaling_factor) as u32,
+    (input_img.height() as f32 * scaling_factor) as u32,
+    imageops::FilterType::Triangle,
+  );
+  Ok(DynamicImage::ImageRgba8(output_img))
+}

crates/desktop/ml_models/src/ml/flux/flux_pipeline.rs

@@ -0,0 +1,297 @@
+use std::path::Path;
+use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+
+use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
+use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
+use crate::ml::model_cache::ModelCache;
+use crate::ml::prompt_cache::PromptCache;
+use candle_nn::VarBuilder;
+use crate::ml::model_config::ModelConfig;
+use anyhow::{anyhow, Error as E, Result};
+use candle_core::{DType, IndexOp, Tensor, D, Device, Module};
+use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
+use image::DynamicImage;
+use log::{debug, info};
+use rand::Rng;
+use tauri::{AppHandle, Emitter};
+use candle_transformers::models::{clip, t5, flux};
+use tokenizers::Tokenizer;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum FluxModel {
+  Schnell,
+  Dev,
+}
+
+pub struct FluxPipelineArgs<'a> {
+  pub image: &'a DynamicImage,
+  pub prompt: String,
+  pub model_cache: &'a ModelCache,
+  pub prompt_cache: &'a PromptCache,
+  pub app: &'a AppHandle,
+  pub flux_model: FluxModel,
+  pub active_blocks: Option<usize>,
+  pub use_cuda_stream: Option<bool>,
+  pub prefetch_next_batch: Option<bool>,
+  pub use_full_gpu: Option<bool>,
+  pub img2img_strength: Option<f64>,
+  pub seed: Option<u64>,
+  pub use_quantized_model: Option<bool>,
+  pub model_config: &'a ModelConfig,
+  pub flux_dev_or_schnell_path: &'a Path,
+  pub flux_schnell_quantized_gguf: &'a Path,
+  pub flux_autoencoder_path: &'a Path,
+  pub google_t5_model_path: &'a Path,
+  pub google_t5_config_path: &'a Path,
+  pub lmz_t5_tokenizer_path: &'a Path,
+  pub openai_clip_model_path: &'a Path,
+  pub openai_clip_tokenizer_json_path: &'a Path,
+}
+
+pub fn flux_pipeline(args: FluxPipelineArgs<'_>) -> Result<RgbImage> {
+  let FluxPipelineArgs { 
+    prompt, 
+    model_cache, 
+    prompt_cache, 
+    app, 
+    image, 
+    flux_model, 
+    active_blocks, 
+    use_cuda_stream, 
+    prefetch_next_batch, 
+    use_full_gpu, 
+    img2img_strength, 
+    seed, 
+    use_quantized_model, 
+    model_config ,
+    flux_dev_or_schnell_path,
+    flux_schnell_quantized_gguf,
+    flux_autoencoder_path,
+    google_t5_model_path,
+    google_t5_config_path,
+    lmz_t5_tokenizer_path,
+    openai_clip_model_path,
+    openai_clip_tokenizer_json_path,
+  } = args;
+
+  if use_quantized_model.unwrap_or(false) {
+    #[cfg(feature = "cuda")]
+    candle_core::quantized::cuda::set_force_dmmv(false);
+  }
+
+  let height = 1024;
+  let width = 1024;
+
+  let quantized = use_quantized_model.unwrap_or(false);
+  let device = model_config.device.clone();
+  let dtype = device.bf16_default_to_f32();
+
+  info!("Starting image generation with FLUX model:");
+  info!("  Model: {:?}", flux_model);
+  info!("  Prompt: {}", prompt);
+  info!("  Dimensions: {}x{}", width, height);
+  info!("  Device: {:?}", model_config.device);
+
+  let seed_value = seed.unwrap_or_else(|| rand::thread_rng().gen());
+  device.set_seed(seed_value)?;
+  
+  info!("Using seed: {}", seed_value);
+
+  // Setup FLUX model configuration
+  let flux_config = match flux_model {
+    FluxModel::Dev => flux::model::Config::dev(),
+    FluxModel::Schnell => flux::model::Config::schnell(),
+  };
+
+  // Get T5 embeddings
+  let t5_emb = {
+    info!("Generating T5 embeddings for prompt...");
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[google_t5_model_path], dtype , &device)? };
+    let config = std::fs::read_to_string(google_t5_config_path)?;
+    let config: t5::Config = serde_json::from_str(&config)?;
+
+    let mut model = t5::T5EncoderModel::load(vb, &config)?;
+    let tokenizer = Tokenizer::from_file(lmz_t5_tokenizer_path).map_err(E::msg)?;
+
+    let mut tokens = tokenizer.encode(prompt.as_str(), true).map_err(E::msg)?.get_ids().to_vec();
+    tokens.resize(256, 0);
+    let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+
+    model.forward(&input_token_ids)?
+  };
+  
+  info!("T5 embeddings generated");
+
+  // Get CLIP embeddings
+  let clip_emb = {
+    info!("Generating CLIP embeddings for prompt...");
+    let vb =
+        unsafe { VarBuilder::from_mmaped_safetensors(&[openai_clip_model_path], dtype, &device)? };
+    // https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+    let config = clip::text_model::ClipTextConfig {
+        vocab_size: 49408,
+        projection_dim: 768,
+        activation: clip::text_model::Activation::QuickGelu,
+        intermediate_size: 3072,
+        embed_dim: 768,
+        max_position_embeddings: 77,
+        pad_with: None,
+        num_hidden_layers: 12,
+        num_attention_heads: 12,
+    };
+    let model =
+        clip::text_model::ClipTextTransformer::new(vb.pp("text_model"), &config)?;
+    let tokenizer = Tokenizer::from_file(openai_clip_tokenizer_json_path).map_err(E::msg)?;
+    let tokens = tokenizer
+        .encode(prompt.as_str(), true)
+        .map_err(E::msg)?
+        .get_ids()
+        .to_vec();
+    let input_token_ids = Tensor::new(&tokens[..], &device)?.unsqueeze(0)?;
+    model.forward(&input_token_ids)?
+  };
+  info!("CLIP embeddings generated");
+
+  // Process input image to tensor
+  info!("Converting input image to tensor...");
+  let img_tensor = dynamic_image_to_tensor(image, &device, dtype)?;
+  info!("Image tensor shape: {:?}", img_tensor.shape());
+
+  // Create initial noise - this is important to match the expected format
+  info!("Generating initial noise for FLUX...");
+  let initial_tensor = flux::sampling::get_noise(1, height, width, &device)?.to_dtype(dtype)?;
+  info!("Initial tensor shape: {:?}", initial_tensor.shape());
+  
+  // Blend initial noise with input image if img2img_strength is specified
+  let initial_tensor = if let Some(strength) = img2img_strength {
+    if strength > 0.0 && strength < 1.0 {
+      info!("Using img2img with strength: {}", strength);
+      // Reshape the input tensor to match the noise tensor dimensions
+      let adjusted_img = if img_tensor.shape() != initial_tensor.shape() {
+        info!("Reshaping input image to match noise dimensions...");
+        // You may need more complex reshaping logic here depending on your use case
+        let (_, c, h, w) = initial_tensor.dims4()?;
+        img_tensor.upsample_nearest2d(h, w)?
+      } else {
+        img_tensor
+      };
+      
+      // Blend the noise and the image according to strength
+      let strength_tensor = Tensor::new(strength as f32, &device)?;
+      let inverse_strength = Tensor::new(1.0 - strength as f32, &device)?;
+      let weighted_noise = (&initial_tensor * &strength_tensor)?;
+      let weighted_img = (&adjusted_img * &inverse_strength)?;
+      (&weighted_noise + &weighted_img)?
+    } else {
+      initial_tensor
+    }
+  } else {
+    initial_tensor
+  };
+  info!("Final initial tensor shape: {:?}", initial_tensor.shape());
+
+  // Create FLUX state
+  info!("Creating FLUX state...");
+  let state = if quantized {
+    flux::sampling::State::new(
+        &t5_emb.to_dtype(DType::F32)?,
+        &clip_emb.to_dtype(DType::F32)?,
+        &initial_tensor.to_dtype(DType::F32)?,
+    )?
+  } else {
+    flux::sampling::State::new(&t5_emb, &clip_emb, &initial_tensor)?
+  };
+  info!("{state:?}");
+
+  // Set appropriate timesteps based on model
+  let timesteps = match flux_model {
+    FluxModel::Dev => flux::sampling::get_schedule(28, Some((state.img.dim(1)?, 0.5, 1.15))),
+    FluxModel::Schnell => flux::sampling::get_schedule(4, None),
+  };
+
+  info!("Starting FLUX denoising with timesteps: {:?}", timesteps);
+
+  // Actual model loading and inference
+  let img_latent = if quantized {
+    // Quantized model path
+    info!("Loading quantized FLUX model...");
+    let model_file = match flux_model {
+      FluxModel::Schnell => flux_schnell_quantized_gguf,
+      FluxModel::Dev => return Err(anyhow!("Quantized model not available for FLUX.1-dev")),
+    };
+
+    info!("Loading quantized model from {}", model_file.to_str().unwrap());
+
+    let vb = candle_transformers::quantized_var_builder::VarBuilder::from_gguf(model_file, &device)?;
+    let mut model = flux::quantized_model::Flux::new(&flux_config, vb)?;
+    info!("Quantized model loaded on {:?}", device);
+
+    // Run denoising
+    info!("Running denoising...");
+    let img_latent = flux::sampling::denoise(&mut model, &state.img, &state.img_ids, &state.txt, &state.txt_ids, &state.vec, &timesteps, 4.)?.to_dtype(dtype)?;
+    info!("Denoising completed");
+    img_latent
+  } else {
+    // Full precision model
+    info!("Loading FLUX model...");
+
+    let mut model = if use_full_gpu.unwrap_or(false) {
+      // Load entire model on GPU
+      info!("Loading entire model directly to GPU...");
+      let gpu_vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_dev_or_schnell_path], dtype, &device)? };
+      let mut model = flux::model::Flux::new(&flux_config, gpu_vb)?;
+      model.set_use_device_management(false);
+      model
+    } else {
+      // Use memory-efficient approach with blocks on CPU
+      info!("Using memory-efficient approach with blocks on CPU...");
+      let cpu_device = candle_core::Device::Cpu;
+      let cpu_vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_dev_or_schnell_path], dtype, &cpu_device)? };
+      let mut model = flux::model::Flux::new_with_gpu_core(&flux_config, cpu_vb, &device)?;
+
+      // Set active blocks
+      let active_blocks_count = active_blocks.unwrap_or(1);
+      model.set_max_active_blocks(active_blocks_count);
+      info!("Using {} active blocks on GPU", active_blocks_count);
+
+      // Enable prefetching if requested
+      if prefetch_next_batch.unwrap_or(false) {
+        model.set_prefetch_next_batch(true);
+        info!("Prefetching enabled - will overlap transfers with computation");
+      }
+
+      model
+    };
+
+    // Run denoising
+    flux::sampling::denoise(&mut model, &state.img, &state.img_ids, &state.txt, &state.txt_ids, &state.vec, &timesteps, 4.)?
+  };
+
+  info!("FLUX denoising completed, unpacking latent image...");
+  let img_unpacked = flux::sampling::unpack(&img_latent, height, width)?;
+
+  // Load autoencoder for final decoding
+  info!("Loading autoencoder for final decoding...");
+  
+  let ae_config = match flux_model {
+    FluxModel::Dev => flux::autoencoder::Config::dev(),
+    FluxModel::Schnell => flux::autoencoder::Config::schnell(),
+  };
+
+  let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[flux_autoencoder_path], dtype, &device)? };
+  let ae_model = flux::autoencoder::AutoEncoder::new(&ae_config, vb)?;
+
+  info!("Decoding final image...");
+  let img = ae_model.decode(&img_unpacked)?;
+
+  // Convert tensor to image
+  info!("Converting tensor to image...");
+  let img = ((img.clamp(-1f32, 1f32)? + 1.0)? * 127.5)?.to_dtype(candle_core::DType::U8)?;
+
+  // Convert to RGB image
+  let image = tensor_to_image_buffer(&img.i(0)?)?;
+
+  info!("FLUX pipeline completed successfully");
+  Ok(image)
+}

crates/desktop/ml_models/src/ml/flux/mod.rs

@@ -0,0 +1 @@
+pub mod flux_pipeline;

crates/desktop/ml_models/src/ml/mod.rs

@@ -1,7 +1,8 @@
 pub mod background_removal;
+pub mod flux;
 pub mod image;
 pub mod model_cache;
+pub mod model_config;
 pub mod model_file;
 pub mod prompt_cache;
-pub mod stable_diffusion;
-pub mod weights_registry;
+pub mod stable_diffusion;

crates/desktop/ml_models/src/ml/model_config.rs

@@ -0,0 +1,61 @@
+use crate::ml::model_file::StableDiffusionVersion;
+use candle_core::{DType, Device};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use hf_hub::api::sync::Api;
+use log::info;
+
+const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
+const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
+const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
+const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::V1_5;
+
+pub struct ModelConfig {
+
+  /// We only support one device type currently, and we can't multiplex
+  /// over several devices yet.
+  pub device: Device,
+
+  /// The core numeric type to use for models.
+  pub dtype: DType,
+
+  pub sd_version: StableDiffusionVersion,
+  pub sd_config: StableDiffusionConfig,
+
+  /// Probably shouldn't be here.
+  pub hf_api: Api,
+}
+
+impl ModelConfig {
+  pub fn init() -> anyhow::Result<Self> {
+    let device = Device::new_cuda(0)
+      .unwrap_or_else(|e| {
+          info!("CUDA not available ({}), falling back to CPU", e);
+          Device::Cpu
+        });
+
+    let sd_config = match SD_VERSION {
+      StableDiffusionVersion::Turbo => {
+        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
+        StableDiffusionConfig::v1_5(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => {
+        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => {
+        StableDiffusionConfig::sdxl(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+    };
+
+    let hf_api = Api::new()?;
+
+    Ok(Self {
+      device,
+      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
+      sd_version: SD_VERSION,
+      sd_config,
+      hf_api,
+    })
+  }
+}

crates/desktop/ml_models/src/ml/stable_diffusion/infer_clip_text_embeddings.rs

@@ -4,18 +4,50 @@ extern crate accelerate_src;
 extern crate intel_mkl_src;
 
 use candle_transformers::models::stable_diffusion;
-use std::path::PathBuf;
+use std::path::{Path, PathBuf};
 
 use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
-use crate::ml::weights_registry::weights::{CLIP_JSON, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2};
-use crate::state::app_dir::AppWeightsDir;
+use ml_weights_registry::weights_registry::weights::{CLIP_JSON, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2};
 use anyhow::Error as E;
 use candle_core::{DType, Device, Module, Tensor, D};
 use log::info;
 use tokenizers::Tokenizer;
 
+
+pub struct ClipArgs<'a, P: AsRef<Path>,  Q: AsRef<Path>> {
+  pub prompt: &'a str, 
+  pub uncond_prompt: &'a str, 
+  pub tokenizer: Option<String>, 
+  pub clip_weights: Option<String>, 
+  pub clip2_weights: Option<String>, 
+  pub sd_version: StableDiffusionVersion, 
+  pub sd_config: &'a stable_diffusion::StableDiffusionConfig, 
+  pub use_f16: bool, 
+  pub device: &'a Device, 
+  pub dtype: DType, 
+  pub use_guide_scale: bool, 
+  pub clip_json_path: P,
+  pub clip_weights_path: Q,
+}
+
 #[allow(clippy::too_many_arguments)]
-pub fn infer_clip_text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer: Option<String>, clip_weights: Option<String>, clip2_weights: Option<String>, sd_version: StableDiffusionVersion, sd_config: &stable_diffusion::StableDiffusionConfig, use_f16: bool, device: &Device, dtype: DType, use_guide_scale: bool, weights_dir: &AppWeightsDir) -> anyhow::Result<Tensor> {
+pub fn infer_clip_text_embeddings<P: AsRef<Path>, Q: AsRef<Path>>(args: ClipArgs<'_, P, Q>) -> anyhow::Result<Tensor> {
+  let ClipArgs {
+    prompt, 
+    uncond_prompt, 
+    tokenizer, 
+    clip_weights, 
+    clip2_weights, 
+    sd_version, 
+    sd_config, 
+    use_f16, 
+    device, 
+    dtype, 
+    use_guide_scale, 
+    clip_json_path,
+    clip_weights_path,
+  } = args;
+
   info!("Preparing text embeddings... for {:?}", sd_version);
 
   let which = match sd_version {
@@ -39,7 +71,8 @@ pub fn infer_clip_text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer:
         dtype,
         use_guide_scale,
         *first,
-        weights_dir,
+        clip_json_path.as_ref(),
+        clip_weights_path.as_ref(),
       )
     })
     .collect::<anyhow::Result<Vec<_>>>()?;
@@ -50,10 +83,25 @@ pub fn infer_clip_text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer:
 }
 
 #[allow(clippy::too_many_arguments)]
-fn do_infer_clip_text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer: Option<String>, clip_weights: Option<String>, clip2_weights: Option<String>, sd_version: StableDiffusionVersion, sd_config: &stable_diffusion::StableDiffusionConfig, use_f16: bool, device: &Device, dtype: DType, use_guide_scale: bool, first: bool, weights_dir: &AppWeightsDir) -> anyhow::Result<Tensor> {
+fn do_infer_clip_text_embeddings(
+  prompt: &str, 
+  uncond_prompt: &str, 
+  tokenizer: Option<String>, 
+  clip_weights: Option<String>, 
+  clip2_weights: Option<String>, 
+  sd_version: StableDiffusionVersion, 
+  sd_config: &stable_diffusion::StableDiffusionConfig, 
+  use_f16: bool, 
+  device: &Device, 
+  dtype: DType, 
+  use_guide_scale: bool, 
+  first: bool, 
+  clip_json_path: &Path,
+  clip_weights_path: &Path,
+) -> anyhow::Result<Tensor> {
   info!("do_infer_clip_text_embeddings called with args {:?}", (prompt, uncond_prompt, tokenizer.clone().unwrap_or_else(|| "None".to_string()), clip_weights.clone().unwrap_or_else(|| "None".to_string()), clip2_weights.clone().unwrap_or_else(|| "None".to_string()), sd_version, sd_config, use_f16, device, dtype, use_guide_scale, first));
 
-  let tokenizer = weights_dir.weight_path(&CLIP_JSON);
+  let tokenizer = clip_json_path;
 
   info!("Loading Clip Tokenizer path: {:?}", tokenizer);
 
@@ -75,11 +123,16 @@ fn do_infer_clip_text_embeddings(prompt: &str, uncond_prompt: &str, tokenizer: O
   let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
 
   println!("Building the Clip transformer.");
+  
+  // TODO(bt,2025-03-23): These used to be two different clip models. We should only need one.
   let clip_weights = if first {
-    weights_dir.weight_path(&SDXL_TURBO_CLIP_TEXT_ENCODER)
+    println!(">>>>> CLIP 1");
+    clip_weights_path
   } else {
-    weights_dir.weight_path(&SDXL_TURBO_CLIP_TEXT_ENCODER_2)
+    println!(">>>>> CLIP 2");
+    clip_weights_path
   };
+  
   let clip_config = if first { &sd_config.clip } else { sd_config.clip2.as_ref().unwrap() };
   let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
 

crates/desktop/ml_models/src/ml/stable_diffusion/lcm_pipeline.rs

@@ -1,53 +1,82 @@
-use std::path::PathBuf;
+use std::path::{Path, PathBuf};
 use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
 
-use crate::events::notification_event::{NotificationEvent, NotificationModelType};
-use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
-use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
-use crate::ml::model_cache::ModelCache;
-use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
-use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
-use crate::ml::weights_registry::weights::{LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
-use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
 use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, IndexOp, Tensor, D};
+use candle_core::{DType, Device, IndexOp, Tensor, D};
 use candle_nn::VarBuilder;
-use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use candle_transformers::models::stable_diffusion::lcm::{LCMScheduler, LCMSchedulerConfig};
 use candle_transformers::models::stable_diffusion::unet_2d::BlockConfig;
 use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
 use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModelConfig;
 use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
+use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
+use crate::ml::model_cache::ModelCache;
+use crate::ml::model_config::ModelConfig;
+use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
+use crate::ml::prompt_cache::PromptCache;
+use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
+use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
 use image::DynamicImage;
-use log::info;
+use log::{debug, error, info};
+use ml_weights_registry::weights_registry::weights::{LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
 use rand::Rng;
-use tauri::{AppHandle, Emitter};
 
 const DEFAULT_STRENGTH : f64 = 75.0;
 
-pub struct Args<'a> {
+const DEFAULT_SEED : u64 = 42;
+
+pub struct Args<'a, P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>> {
   pub image: &'a DynamicImage,
   pub prompt: String,
   pub uncond_prompt: String,
   pub cfg_scale: Option<f64>,
-  pub i2i_strength: Option<f64>,
-  pub configs: &'a AppConfig,
+  pub img2img_strength: Option<f64>,
   pub model_cache: &'a ModelCache,
   pub prompt_cache: &'a PromptCache,
-  pub app_data_root: &'a AppDataRoot,
-  pub app: &'a AppHandle,
   pub use_flash_attn: bool,
+  pub model_config: &'a ModelConfig,
+  pub maybe_seed: Option<u64>,
+  pub scheduler_steps: usize,
+  pub vae_path: P1,
+  pub unet_path: P2,
+  pub clip_json_path: P3,
+  pub clip_weights_path: P4,
 }
 
-pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
-  let Args { prompt, uncond_prompt, cfg_scale, i2i_strength, configs, model_cache, prompt_cache, app, image, app_data_root, use_flash_attn } = args;
-
-  let weights_dir = app_data_root.weights_dir();
+pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
+  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
+{
+  let Args { 
+    prompt, 
+    uncond_prompt, 
+    cfg_scale,
+    img2img_strength, 
+    model_cache, 
+    prompt_cache, 
+    image, 
+    use_flash_attn ,
+    model_config: ModelConfig {
+      device,
+      dtype,
+      sd_version,
+      sd_config,
+      hf_api,
+    },
+    maybe_seed,
+    scheduler_steps,
+    vae_path,
+    unet_path,
+    clip_json_path,
+    clip_weights_path,
+  } = args;
   
-  let img2img_strength = args.i2i_strength
+  let dtype = *dtype;
+  let sd_version = *sd_version;
+
+  let img2img_strength = img2img_strength
     .map(|s| remap_lcm_strength_range(s))
     .unwrap_or(DEFAULT_STRENGTH);
   
@@ -56,14 +85,16 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
   let use_f16 = true;
 
   // Use LCM Scheduler instead of Euler Ancestral for better speed and quality
-  let mut scheduler = LCMScheduler::new(configs.scheduler_steps, img2img_strength, candle_transformers::models::stable_diffusion::lcm::LCMSchedulerConfig::default())?;
+  let mut scheduler = LCMScheduler::new(scheduler_steps, img2img_strength, LCMSchedulerConfig::default())?;
 
-  let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
-  configs.device.set_seed(seed)?;
+  //let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
+  let seed = maybe_seed.unwrap_or(DEFAULT_SEED);
+
+  device.set_seed(seed)?;
 
   let guidance_scale = match cfg_scale {
     Some(guidance_scale) => guidance_scale,
-    None => match configs.sd_version {
+    None => match sd_version {
       StableDiffusionVersion::V1_5 | StableDiffusionVersion::V2_1 | StableDiffusionVersion::Xl => 7.5,
       StableDiffusionVersion::Turbo => 0.,
       _ => 0., // NB: Not sure what the other model families should use, so sticking with "0"
@@ -78,50 +109,45 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
   let mut text_embeddings = if let Some(tensor) = maybe_cached {
     tensor
   } else {
-    info!("Loading clip weights...");
-
-    let clip_weights = weights_dir.weight_path(&LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16);
-    let clip_weights = Some(clip_weights.to_string_lossy().to_string());
-
     info!("Prompt is NOT cached! Calculating embedding...");
-    let tensor = infer_clip_text_embeddings(
-      &prompt,
-      &uncond_prompt,
-      None, // tokenizer
-      clip_weights,
-      None, // clip2_weights
-      configs.sd_version,
-      &configs.sd_config,
+    let tensor = infer_clip_text_embeddings(ClipArgs {
+      prompt: & prompt,
+      uncond_prompt: &uncond_prompt,
+      tokenizer: None, // tokenizer
+      clip_weights: None,
+      clip2_weights: None, // clip2_weights
+      sd_version,
+      sd_config: &sd_config,
       use_f16, // use_f16
-      &configs.device,
-      configs.dtype,
+      device,
+      dtype,
       use_guide_scale,
-      weights_dir,
-    )?;
+      clip_json_path,
+      clip_weights_path,
+    })?;
+
     prompt_cache.store_copy(&prompt, &tensor)?;
     tensor
   };
 
-  println!("Text embeddings shape: {:?}", text_embeddings.shape());
+  debug!("Text embeddings shape: {:?}", text_embeddings.shape());
 
-  println!("Loading input image into tensor...");
+  debug!("Loading input image into tensor...");
 
-  let input_image = dynamic_image_to_tensor(image, &configs.device, configs.dtype)?;
+  let input_image = dynamic_image_to_tensor(image, &device, dtype)?;
 
-  println!("Reference image shape: {:?}", input_image.shape());
+  debug!("Reference image shape: {:?}", input_image.shape());
 
-  let vae_scale = get_vae_scale(configs.sd_version);
+  let vae_scale = get_vae_scale(sd_version);
 
   let maybe_vae = model_cache.get_vae()?;
 
   let vae = match maybe_vae {
     Some(vae) => vae,
     None => {
-      let vae_file = weights_dir.weight_path(&LYKON_DEAMSHAPER_7_VAE);
-
-      println!("Building VAE model from file {:?}...", &vae_file);
-
-      let vae = configs.sd_config.build_vae(vae_file, &configs.device, configs.dtype)?;
+      info!("Building VAE model from file {:?}...", vae_path.as_ref());
+      
+      let vae = sd_config.build_vae(vae_path, &device, dtype)?;
 
       let vae = Arc::new(vae);
 
@@ -136,11 +162,9 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
   let unet = match maybe_unet {
     Some(unet) => unet,
     None => {
-      info!("No unet found in cache; loading...");
+      info!("Building UNET model from file {:?}...", unet_path.as_ref());
 
-      let unet_weights = weights_dir.weight_path(&SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
-
-      let in_channels = match configs.sd_version {
+      let in_channels = match sd_version {
         StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
         _ => 4,
       };
@@ -175,7 +199,7 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
       };
 
       // Load the model directly from the safetensors file
-      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_weights], configs.dtype, &configs.device)? };
+      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_path], dtype, &device)? };
 
       let unet = UNet2DConditionModel::new(vs_unet, in_channels, 4, use_flash_attn, unet_config)?;
 
@@ -191,19 +215,20 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
   // Create LCM guidance scale embeddings
   let embedding_dim = 1280;
-  let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &configs.device, configs.dtype)?;
+  let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &device, dtype)?;
 
   // Generate latents from input image using the LCM approach
-  println!("Generating latents from input image...");
+  debug!("Generating latents from input image...");
   let init_latents = (init_latent_dist.sample()? * vae_scale)?;
-  println!("Initial latents shape: {:?}", init_latents.shape());
+
+  debug!("Initial latents shape: {:?}", init_latents.shape());
 
   // Calculate timesteps for LCM with proper img2img handling
-  let timesteps = LCMScheduler::get_timesteps_for_steps(configs.scheduler_steps, img2img_strength);
+  let timesteps = LCMScheduler::get_timesteps_for_steps(scheduler_steps, img2img_strength);
   let t_start = timesteps[0]; // First denoising step
 
   // Add noise at the appropriate timestep for img2img
-  println!("Adding noise to latents for t_start={}", t_start);
+  debug!("Adding noise to latents for t_start={}", t_start);
   let noise = init_latents.randn_like(0f64, 1f64)?;
   let latents = scheduler.add_noise(&init_latents, noise, t_start)?;
 
@@ -211,7 +236,7 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
   let timesteps: Vec<_> = scheduler.timesteps().iter().copied().collect();
 
   for (timestep_index, &timestep) in timesteps.iter().enumerate() {
-    println!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
+    info!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
 
     let latent_model_input = if use_guide_scale { Tensor::cat(&[&latents, &latents], 0)? } else { latents.clone() };
 
@@ -221,7 +246,7 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
     let mut noise_pred = match unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, Some(&guidance_scale_embedding)) {
       Ok(pred) => pred,
       Err(e) => {
-        println!("UNet inference failed with error: {}", e);
+        error!("UNet inference failed with error: {}", e);
         return Err(anyhow::anyhow!("UNet inference failed: {}", e));
       },
     };
@@ -233,24 +258,22 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
     }
 
     // Apply the LCM scheduler step
-    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, configs.scheduler_steps)?;
+    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, scheduler_steps)?;
   }
 
-  println!("Diffusion process completed, decoding image...");
+  debug!("Diffusion process completed, decoding image...");
   let image = vae.decode(&(latents / vae_scale)?)?;
 
-  println!("VAE decode completed");
-
-  println!("Scaling image back");
+  debug!("Scaling image back");
   let image = ((image / 2.)? + 0.5)?;
 
-  println!("Normalized image values");
+  debug!("Normalized image values");
   let image = (image.clamp(0f32, 1.)? * 255.)?;
 
-  println!("Convert to int8");
+  debug!("Convert to int8");
   let image = image.to_dtype(DType::U8)?;
 
-  println!("Converted to 8-bit format");
+  debug!("Converted to 8-bit format");
 
   let image = tensor_to_image_buffer(&image.i(0)?)?;
 

crates/desktop/ml_models/src/ml/stable_diffusion/stable_diffusion_pipeline.rs

@@ -1,75 +1,78 @@
+use std::path::Path;
 use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
 
-use crate::events::notification_event::{NotificationEvent, NotificationModelType};
 use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
 use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
 use crate::ml::model_cache::ModelCache;
 use crate::ml::model_file::StableDiffusionVersion;
 use crate::ml::prompt_cache::PromptCache;
 use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
-use crate::ml::weights_registry::weights::{SDXL_TURBO_UNET, SDXL_TURBO_VAE};
-use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
 use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, IndexOp, Tensor, D};
+use candle_core::{DType, Device, IndexOp, Tensor, D};
 use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
 use image::DynamicImage;
 use log::info;
 use rand::Rng;
-use tauri::{AppHandle, Emitter};
 
-pub struct Args<'a> {
+pub struct Args<'a, P: AsRef<Path>, Q: AsRef<Path>, R: AsRef<Path>, S: AsRef<Path>> {
     pub image: &'a DynamicImage,
     pub prompt: String,
     pub uncond_prompt: String,
     pub cfg_scale: Option<f64>,
     pub i2i_strength: Option<u8>,
-    pub configs: &'a AppConfig,
     pub model_cache: &'a ModelCache,
     pub prompt_cache: &'a PromptCache,
-    pub app_data_root: &'a AppDataRoot,
-    pub app: &'a AppHandle,
+    pub sd_version: StableDiffusionVersion,
+    pub sd_config: StableDiffusionConfig,
+    pub device: &'a Device,
+    pub dtype: DType,
+    pub maybe_seed: Option<u64>,
+    pub scheduler_steps: usize,
+    pub sdxl_turbo_vae_path: P,
+    pub sdxl_turbo_unet_path: Q,
+    pub clip_json_path: R,
+    pub clip_weights_path: S,
 }
 
-pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
+pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
+  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
+{
     let Args { 
         prompt, 
         uncond_prompt, 
         cfg_scale, 
         i2i_strength, 
-        configs, 
         model_cache, 
         prompt_cache, 
-        app, 
         image,
-        app_data_root,
+        sd_version,
+        sd_config,
+        device,
+        dtype,
+        maybe_seed,
+        scheduler_steps,
+        sdxl_turbo_vae_path,
+        sdxl_turbo_unet_path,
+        clip_json_path,
+        clip_weights_path,
     } = args;
     
-    let weights_dir = app_data_root.weights_dir();
-    
-    println!("Starting image generation with the following configuration:");
-    println!("  Model: {:?}", configs.sd_version);
-    println!("  Prompt: {}", prompt);
-    println!("  Steps: {}", configs.scheduler_steps);
-    println!("  Device: {:?}", configs.device);
-
-    println!("Model dimensions: {}x{}", configs.sd_config.width, configs.sd_config.height);
-
     // TODO(bt,2025-02-18): The scheduler is `EulerAncestralDiscreteScheduler`, but we may want to port an LCM scheduler.
     //  This is a target for performance improvement
     //  See: https://github.com/huggingface/candle/issues/1331
-    let mut scheduler = configs.sd_config.build_scheduler(
-        configs.scheduler_steps)?;
+    let mut scheduler = sd_config.build_scheduler(scheduler_steps)?;
 
-    let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
-    configs.device.set_seed(seed)?;
+    let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
+    
+    device.set_seed(seed)?;
 
     info!("Using seed: {}", seed);
 
     let guidance_scale = match cfg_scale {
         Some(guidance_scale) => guidance_scale,
-        None => match configs.sd_version {
+        None => match sd_version {
             StableDiffusionVersion::V1_5
             | StableDiffusionVersion::V2_1
             | StableDiffusionVersion::Xl => 7.5,
@@ -90,20 +93,21 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
     } else {
         
         info!("Prompt is NOT cached! Calculating embedding...");
-        let tensor = infer_clip_text_embeddings(
-            &prompt,
-            &uncond_prompt,
-            None, // tokenizer
-            None, // clip_weights
-            None, // clip2_weights
-            configs.sd_version,
-            &configs.sd_config,
-            false, // use_f16
-            &configs.device,
-            configs.dtype,
+        let tensor = infer_clip_text_embeddings(ClipArgs {
+            prompt: &prompt,
+            uncond_prompt: &uncond_prompt,
+            tokenizer: None, // tokenizer
+            clip_weights: None, // clip_weights
+            clip2_weights: None, // clip2_weights
+            sd_version: sd_version,
+            sd_config: &sd_config,
+            use_f16: false, // use_f16
+            device: device,
+            dtype: dtype,
             use_guide_scale,
-            weights_dir,
-        )?;
+            clip_json_path: clip_json_path.as_ref(),
+            clip_weights_path: clip_weights_path.as_ref(),
+        })?;
         prompt_cache.store_copy(&prompt, &tensor)?;
         tensor
     };
@@ -112,14 +116,11 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
     println!("Loading input image into tensor...");
 
-    let input_image = dynamic_image_to_tensor(
-        image,
-        &configs.device,
-        configs.dtype)?;
+    let input_image = dynamic_image_to_tensor(image, device, dtype)?;
 
     println!("Reference image shape: {:?}", input_image.shape());
 
-    let vae_scale = get_vae_scale(configs.sd_version);
+    let vae_scale = get_vae_scale(sd_version);
 
     let maybe_vae = model_cache.get_vae()?;
     
@@ -128,11 +129,9 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No vae found in cache; loading...");
             
-            let vae_file = weights_dir.weight_path(&SDXL_TURBO_VAE);
+            let vae_file = sdxl_turbo_vae_path.as_ref();
             
-            let vae = configs
-              .sd_config
-              .build_vae(vae_file, &configs.device, configs.dtype)?;
+            let vae = sd_config.build_vae(vae_file, &device, dtype)?;
 
             let vae = Arc::new(vae);
             
@@ -149,9 +148,9 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No unet found in cache; loading...");
 
-            let unet_weights_file = weights_dir.weight_path(&SDXL_TURBO_UNET);
+            let unet_weights_file = sdxl_turbo_unet_path.as_ref();
 
-            let unet = configs.sd_config.build_unet(unet_weights_file, &configs.device, 4, true, configs.dtype)
+            let unet = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)
               .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
 
             let unet = Arc::new(unet);
@@ -181,9 +180,9 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
     };
 
     let t_start = {
-        let start = configs.scheduler_steps - (configs.scheduler_steps as f64 * img2img_strength) as usize;
+        let start = scheduler_steps - (scheduler_steps as f64 * img2img_strength) as usize;
 
-        println!("Starting from step {} of {} (strength: {})", start, configs.scheduler_steps, img2img_strength);
+        println!("Starting from step {} of {} (strength: {})", start, scheduler_steps, img2img_strength);
         start
     };
 

crates/desktop/ml_weights_registry/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "ml_weights_registry"
+edition = "2021"
+version = "0.0.1"
+authors = ["storyteller"]
+publish = false
+description = ""
+license = ""
+repository = ""
+
+[lib]
+name = "ml_weights_registry"
+path = "src/lib.rs"
+
+# External
+[dependencies]
+anyhow = "1.0"
+const_format = "0.2.34"

crates/desktop/ml_weights_registry/src/lib.rs

@@ -0,0 +1 @@
+pub mod weights_registry;

crates/desktop/ml_weights_registry/src/weights_registry/flux_weights.rs

@@ -0,0 +1,80 @@
+use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
+use crate::weights_registry::weight_descriptor_builder::weight;
+
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-dev
+pub const FLUX_DEV : WeightDescriptor = weight!(
+  "Flux-1 Dev",
+  "flux1-dev.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+// TODO(bt): I think the autoencoder for Dev and Schnell are the same.
+//  They have the same checksum unless I mixed up the files when downloading.
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-dev
+pub const FLUX_DEV_AUTOENCODER : WeightDescriptor = weight!(
+  "Flux-1 Dev (Autoencoder)",
+  "flux1-dev.ae.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-schnell
+pub const FLUX_SCHNELL : WeightDescriptor = weight!(
+  "Flux-1 Schnell",
+  "flux1-schnell.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/black-forest-labs/FLUX.1-schnell
+pub const FLUX_SCHNELL_AUTOENCODER : WeightDescriptor = weight!(
+  "Flux-1 Schnell (Autoencoder)",
+  "flux1-schnell.ae.safetensors",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/lmz/candle-flux
+pub const LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF : WeightDescriptor = weight!(
+  "LMZ Flux Schnell Quantized",
+  "lmz_candle_flux_flux1-schnell.gguf",
+  WeightFunction::ImageGeneration
+);
+
+/// https://huggingface.co/google/t5-v1_1-xxl/tree/main
+/// NB: I couldn't find the safetensors file here
+/// NB: This should live under "refs/pr/2"
+pub const GOOGLE_T5_V1_1_XXL_MODEL : WeightDescriptor = weight!(
+  "Google T5 v1.1 (XXL)",
+  "google_t5_v1_1_xxl_model.safetensors",
+  WeightFunction::TextEncoder
+);
+
+/// https://huggingface.co/google/t5-v1_1-xxl/tree/main
+/// NB: I couldn't find the safetensors file here
+/// NB: This should live under "refs/pr/2"
+pub const GOOGLE_T5_V1_1_XXL_CONFIG : WeightDescriptor = weight!(
+  "Google T5 v1.1 (XXL) Config",
+  "google_t5_v1_1_xxl_config.json",
+  WeightFunction::TextEncoder
+);
+
+/// https://huggingface.co/lmz/mt5-tokenizers
+pub const LMZ_T5_TOKENIZER_JSON : WeightDescriptor = weight!(
+  "LMZ T5 Tokenizer",
+  "lmz_mt5_tokenizers_t5-v1_1-xxl.tokenizer.json",
+  WeightFunction::TextTokenizer
+);
+
+/// https://huggingface.co/openai/clip-vit-large-patch14/tree/main
+pub const OPENAI_CLIP_VIT_P14_MODEL : WeightDescriptor = weight!(
+  "OpenAI CLIP VIT (patch 14) model",
+  "openai_clip-vit-large-patch14.model.safetensors",
+  WeightFunction::TextTokenizer
+);
+
+/// https://huggingface.co/openai/clip-vit-large-patch14/tree/main
+pub const OPENAI_CLIP_VIT_P14_TOKENIZER_JSON: WeightDescriptor = weight!(
+  "OpenAI CLIP VIT (patch 14) tokenizer",
+  "openai_clip-vit-large-patch14.tokenizer.json",
+  WeightFunction::TextTokenizer
+);

crates/desktop/ml_weights_registry/src/weights_registry/mod.rs

@@ -1,3 +1,4 @@
 mod weight_descriptor_builder;
+pub mod flux_weights;
 pub mod weight_descriptor;
 pub mod weights;

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor.rs

@@ -5,6 +5,7 @@
 
 #[derive(Clone, Copy)]
 pub enum WeightFunction {
+  ImageGeneration,
   ImageSegmentation,
   TextEncoder,
   TextTokenizer,
@@ -29,6 +30,7 @@ pub struct WeightDescriptor {
 impl WeightFunction {
   pub fn as_descriptive_str(&self) -> &'static str {
     match self {
+      WeightFunction::ImageGeneration => "image generation",
       WeightFunction::ImageSegmentation => "image segmentation",
       WeightFunction::TextEncoder => "text encoder",
       WeightFunction::TextTokenizer => "text tokenizer",

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor_builder.rs

@@ -5,10 +5,10 @@ pub (super) const R2_BUCKET_URL : &str = "https://pub-bc5e2bc0cdee4bb5ae8fca9d64
 //  and variables (even if they're known to be &'static) and keep allocations static.
 macro_rules! weight {
     ($name: literal, $file_name:literal, $weight_function:expr) => {
-      $crate::ml::weights_registry::weight_descriptor::WeightDescriptor {
+      $crate::weights_registry::weight_descriptor::WeightDescriptor {
         name: $name,
         filename: $file_name,
-        r2_download_url: const_format::concatcp!(crate::ml::weights_registry::weight_descriptor_builder::R2_BUCKET_URL, $file_name),
+        r2_download_url: const_format::concatcp!(crate::weights_registry::weight_descriptor_builder::R2_BUCKET_URL, $file_name),
         function: $weight_function,
       }
     };

crates/desktop/ml_weights_registry/src/weights_registry/weights.rs

@@ -1,5 +1,5 @@
-use crate::ml::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
-use crate::ml::weights_registry::weight_descriptor_builder::weight;
+use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
+use crate::weights_registry::weight_descriptor_builder::weight;
 
 pub const CLIP_JSON : WeightDescriptor = weight!(
   "CLIP",

crates/desktop/src-tauri/Cargo.toml

@@ -7,6 +7,7 @@ license = ""
 repository = ""
 edition = "2021"
 rust-version = "1.77.2"
+publish = false
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
@@ -20,35 +21,35 @@ tauri-build = { version = "2.0.4", features = [] }
 [dependencies]
 
 # Internal
+crockford = { path = "../../lib/data/crockford" }
 filesys = { path = "../../lib/files/filesys" }
 memory_caching = { path = "../../lib/caching/memory_caching" }
+ml_weights_registry = { path = "../ml_weights_registry" }
 primitives = { path = "../../lib/data/primitives" }
 
+# Internal ML
+# TODO: Only on Windows and Linux
+#ml_models = { path = "../ml_models", optional = true }
 # Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
-cudarc = "0.13.7"
+cudarc = { version = "0.13.5", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
 
 # External
 #openssl = { version = "0.10", features = ["vendored"] }
+# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
 anyhow = "1.0"
 base64 = "0.22.1"
 bytes = "1.10"
-candle-core = { path = "../../vendor/candle/candle-core", features = ["cuda"] }
-candle-nn = { path = "../../vendor/candle/candle-nn", features = ["cuda"] }
-candle-transformers = { path = "../../vendor/candle/candle-transformers", features = ["cuda"] }
+chrono = "0.4.40"
 const_format = "0.2.34"
 directories = "6.0.0"
 dirs = "6.0.0"
 futures = "0.3"
 half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-hf-hub = "0.3.2"
 image = "0.25.5"
 log = "0.4"
-ndarray = "0.15.6"
-nvidia-checker = "0.1.2"
 once_cell = "1.20.3"
-ort = { version = "1.16.2", features = ["cuda"] }
+ort = { version = "1.16.2", features = ["cuda", "load-dynamic"] }
 rand = "0.8.5"
-# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
 reqwest = { version = "0.12.12", features = ["stream"], default-features = false }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
@@ -57,28 +58,23 @@ tauri = { version = "2.2.4", features = [] }
 tauri-plugin-log = "2"
 tempdir = "0.3.7"
 tempfile = "3.18.0"
-tokenizers = "0.15.2"
 tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
 tokio-util = { version = "0.7", features = ["codec"] }
 
 # tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
 
-[features]
-default = ["cuda"]
-cuda = [
-    "candle-core/cuda",
-    "candle-nn/cuda",
-    "candle-transformers/cuda",
-# NB(bt,2025-03-10): Building without Flash attention on Windows for now to bypass linker issues
-    "candle-transformers/flash-attn",
-    "candle-core/cudnn"
-]
+[target.'cfg(not(target_os = "macos"))'.dependencies]
+ml_models = { path = "../ml_models" }
 
 [dev-dependencies]
 libloading = "0.8.6"
 serial_test = "3.2.0"
 assertor = "0.0.3"
 
+#[target.'cfg(target_os = "macos")'.features]
+#default  = []
+#ml_models = ["dep:ml_models"]
+
 [profile.dev]
 incremental = true # Compile your binary in smaller steps.
 

crates/desktop/src-tauri/src/endpoints/inpaint_image_endpoint.rs

@@ -0,0 +1,13 @@
+use log::info;
+
+
+#[tauri::command]
+pub async fn inpaint_image(
+  image: &str,
+  mask: &str,
+  prompt: String,
+) -> Result<String, String> {
+  info!("inpaint_endpoint endpoint called.");
+
+  Ok(mask.to_string())
+}

crates/desktop/src-tauri/src/endpoints/mod.rs

@@ -1,6 +1,7 @@
 pub mod download_models;
 pub mod flip_image;
+pub mod inpaint_image_endpoint;
 pub mod realtime_image_endpoint;
 pub mod remove_background_endpoint;
-pub mod test_counter;
+pub mod save_image_endpoint;
 pub mod text_to_image_endpoint;

crates/desktop/src-tauri/src/endpoints/realtime_image_endpoint.rs

@@ -1,16 +1,21 @@
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
+#[cfg(not(target_os = "macos"))]
+use ml_models::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
+
 use crate::state::app_config::AppConfig;
 use crate::state::app_dir::AppDataRoot;
+use crate::stubs::model_cache::ModelCache;
+use crate::stubs::prompt_cache::PromptCache;
 use crate::utils::image::decode_base64_image::decode_base64_image;
+use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
 use crate::utils::image::encode_rgb_image_base64_png::encode_rgb_image_base64_png;
 use image::imageops::FilterType;
 use image::{DynamicImage, RgbImage};
 use log::{error, info};
+use ml_weights_registry::weights_registry::weights::{CLIP_JSON, LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
 use std::path::PathBuf;
 use tauri::{AppHandle, State};
 
+
 const PROMPT_FILENAME : &str = "prompt.txt";
 const PROMPT: &str = "shiba inu, cute dog, detailed, walking in a wooded forest, photorealistic, 8k";
 
@@ -26,6 +31,7 @@ pub async fn infer_image(
   image: &str,
   prompt: Option<String>,
   strength: Option<f64>,
+  seed: Option<u64>,
   model_config: State<'_, AppConfig>,
   model_cache: State<'_, ModelCache>,
   prompt_cache: State<'_, PromptCache>,
@@ -42,28 +48,33 @@ pub async fn infer_image(
     .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
   
   // TODO(bt,2025-02-17): Running out of vram with full image buffer size
-  let image = image.resize(512, 512, FilterType::CatmullRom);
+  let mut image = image.resize(512, 512, FilterType::CatmullRom);
 
-  let result = do_infer_image(
-    &prompt, 
-    image, 
-    strength, 
-    &model_config, 
-    &model_cache, 
-    prompt_cache, 
-    app, 
-    &app_data_root
-  ).await;
-  
-  let image = match result {
-    Ok(image) => image,
-    Err(err) => {
-      error!("There was an error generating the image: {:?}", err);
-      return Err(format!("There was an error generating the image: {}", err));
-    }
-  };
+  let seed = seed.or(model_config.seed);
 
-  let bytes = encode_rgb_image_base64_png(image)
+  #[cfg(not(target_os = "macos"))]
+  {
+    let result = infer_image_impl(
+      &prompt,
+      image,
+      strength,
+      seed,
+      &model_config,
+      &model_cache,
+      prompt_cache,
+      &app_data_root
+    ).await;
+
+    image = match result {
+      Ok(img) => DynamicImage::ImageRgb8(img),
+      Err(err) => {
+        error!("There was an error generating the image: {:?}", err);
+        return Err(format!("There was an error generating the image: {}", err));
+      }
+    };
+  }
+
+  let bytes = encode_dynamic_image_base64_png(image)
     .map_err(|err| format!("failure to encode image: {:?}", err))?;
 
   info!("Inference successful; image converted to base64, serving back to browser...");
@@ -71,29 +82,41 @@ pub async fn infer_image(
   Ok(bytes)
 }
 
-async fn do_infer_image(
+#[cfg(not(target_os = "macos"))]
+async fn infer_image_impl(
   prompt: &str,
   image: DynamicImage,
   strength: Option<f64>,
+  seed: Option<u64>,
   config: &AppConfig,
   model_cache: &ModelCache,
   prompt_cache: State<'_, PromptCache>,
-  app: AppHandle,
   app_data_root: &AppDataRoot,
 ) -> Result<RgbImage, String> {
+  
+  let weights_dir = app_data_root.weights_dir();
+  let vae_path = weights_dir.weight_path(&LYKON_DEAMSHAPER_7_VAE);
+  let unet_path = weights_dir.weight_path(&SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
+  let clip_json_path = weights_dir.weight_path(&CLIP_JSON);
+  //let clip_weights_path= weights_dir.weight_path(&LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16);
+  let clip_weights_path= weights_dir.weight_path(&SDXL_TURBO_CLIP_TEXT_ENCODER);
 
   let args = Args {
     image: &image,
     prompt: prompt.to_string(),
     uncond_prompt: NEGATIVE_PROMPT.to_string(),
     model_cache,
-    configs: config,
     prompt_cache: &prompt_cache,
-    i2i_strength: strength,
+    img2img_strength: strength,
     cfg_scale: config.cfg_scale,
-    app: &app,
-    app_data_root,
     use_flash_attn: true,
+    model_config: &config.model_config,
+    maybe_seed: seed,
+    scheduler_steps: config.scheduler_steps,
+    vae_path: &vae_path,
+    unet_path: &unet_path,
+    clip_json_path: &clip_json_path,
+    clip_weights_path: &clip_weights_path,
   };
 
   let image = lcm_pipeline(args)

crates/desktop/src-tauri/src/endpoints/remove_background_endpoint.rs

@@ -1,43 +1,31 @@
 //! Adapted from https://github.com/dnanhkhoa/rust-background-removal
 
-use crate::ml::background_removal::onnx_session::onnx_session;
-use crate::ml::model_cache::ModelCache;
+#[cfg(not(target_os = "macos"))]
+use ml_models::ml::background_removal::remove_image_background::remove_image_background;
+
 use crate::state::app_dir::AppDataRoot;
 use crate::utils::image::decode_base64_image::decode_base64_image;
 use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
 use image::{imageops, DynamicImage};
-use log::{error, info};
-use ndarray::{Array, CowArray};
-use ort::Value;
-use tauri::{AppHandle, State};
-use crate::ml::weights_registry::weights::{DIS_MEDIUM_ONNX, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
+use log::info;
+use ml_weights_registry::weights_registry::weights::DIS_MEDIUM_ONNX;
+use tauri::State;
 
 /// This handler removes the background from an image.
 #[tauri::command]
 pub async fn remove_background(
   image: &str,
-  model_cache: State<'_, ModelCache>,
   app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
 ) -> Result<String, String> {
   info!("remove_background endpoint called.");
 
-  let image = decode_base64_image(image)
-    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
+  let mut image = decode_base64_image(image)
+      .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
 
-  // NB: Real time needed cuDNN 9, but this uses cuDNN 8:
-  // sudo apt-get install libcudnn8-dev libcudnn8
-  // update-alternatives: warning: forcing reinstallation of alternative /usr/include/x86_64-linux-gnu/cudnn_v9.h because link group libcudnn is broken
-  // update-alternatives: using /usr/include/x86_64-linux-gnu/cudnn_v8.h to provide /usr/include/cudnn.h (libcudnn) in manual mode
-  let model_path = app_data_root.weights_dir().weight_path(&DIS_MEDIUM_ONNX);
-  let session = onnx_session(model_path.to_str().unwrap()) // TODO
-    .map_err(|err| format!("failure to start onnx session: {:?}", err))?;
-
-  let image = process_dynamic_image(&session, image)
-    .map_err(|err| {
-      error!("Couldn't remove background from image: {}", err);
-      format!("Couldn't remove background from image: {}", err)
-    })?;
+  #[cfg(not(target_os = "macos"))]
+  {
+    image = remove_background_impl(&app_data_root, image).await?;
+  }
 
   let image = encode_dynamic_image_base64_png(image)
     .map_err(|err| format!("failure to encode image: {:?}", err))?;
@@ -45,48 +33,13 @@ pub async fn remove_background(
   Ok(image)
 }
 
-fn process_dynamic_image(
-  session: &ort::Session,
-  dynamic_img: DynamicImage,
-) -> Result<DynamicImage, anyhow::Error> {
-  let input_shape = session.inputs[0]
-    .dimensions()
-    .map(|dim| dim.unwrap())
-    .collect::<Vec<usize>>();
-  let input_img = dynamic_img.into_rgba8();
-  let scaling_factor = f32::min(
-    1., // Avoid upscaling
-    f32::min(
-      input_shape[3] as f32 / input_img.width() as f32, // Width ratio
-      input_shape[2] as f32 / input_img.height() as f32, // Height ratio
-    ),
-  );
-  let mut resized_img = imageops::resize(
-    &input_img,
-    input_shape[3] as u32,
-    input_shape[2] as u32,
-    imageops::FilterType::Triangle,
-  );
-  let input_tensor = CowArray::from(
-    Array::from_shape_fn(input_shape, |indices| {
-      let mean = 128.;
-      let std = 256.;
+#[cfg(not(target_os = "macos"))]
+async fn remove_background_impl(app_data_root: &AppDataRoot, image: DynamicImage) -> Result<DynamicImage, String> {
+  let model_path = app_data_root.weights_dir().weight_path(&DIS_MEDIUM_ONNX);
 
-      (resized_img[(indices[3] as u32, indices[2] as u32)][indices[1]] as f32 - mean) / std
-    })
-      .into_dyn(),
-  );
-  let inputs = vec![Value::from_array(session.allocator(), &input_tensor)?];
-  let outputs = session.run(inputs)?;
-  let output_tensor = outputs[0].try_extract::<f32>()?;
-  for (indices, alpha) in output_tensor.view().indexed_iter() {
-    resized_img[(indices[3] as u32, indices[2] as u32)][3] = (alpha * 255.) as u8;
-  }
-  let output_img = imageops::resize(
-    &resized_img,
-    (input_img.width() as f32 * scaling_factor) as u32,
-    (input_img.height() as f32 * scaling_factor) as u32,
-    imageops::FilterType::Triangle,
-  );
-  Ok(DynamicImage::ImageRgba8(output_img))
-}
+  let image = remove_image_background(model_path, image)
+      .await
+      .map_err(|err| format!("Couldn't remove image from image: {}", err))?;
+
+  Ok(image)
+}

crates/desktop/src-tauri/src/endpoints/save_image_endpoint.rs

@@ -0,0 +1,38 @@
+use crate::state::app_config::AppConfig;
+use crate::state::app_dir::AppDataRoot;
+use crate::stubs::model_cache::ModelCache;
+use crate::stubs::prompt_cache::PromptCache;
+use crate::utils::image::decode_base64_image::decode_base64_image;
+use chrono::Local;
+use crockford::crockford_entropy_lower;
+use image::{ImageFormat, RgbImage};
+use log::info;
+use tauri::{AppHandle, State};
+
+const FILENAME_LENGTH : usize = 10;
+
+/// This handler takes an image (as a base64 encoded string) and saves it.
+#[tauri::command]
+pub async fn save_image(
+  image: &str,
+  app_data_root: State<'_, AppDataRoot>,
+) -> Result<String, String> {
+  info!("save_image endpoint called.");
+
+  let image = decode_base64_image(image)
+    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
+
+  let assets_dir = app_data_root.assets_dir();
+  
+  let date_save_dir = assets_dir.make_or_get_current_date_dir()
+    .map_err(|err| format!("Couldn't create assets save directory: {}", err))?;
+
+  let entropy = crockford_entropy_lower(FILENAME_LENGTH); // TODO: These are ugly. A monotonic counter would be better.
+  let filename = format!("{}.png", entropy);
+  let filename = date_save_dir.join(filename);
+  
+  image.save_with_format(filename,  ImageFormat::Png)
+    .map_err(|err| format!("Couldn't save image: {}", err))?;
+
+  Ok("".to_string())
+}

crates/desktop/src-tauri/src/endpoints/test_counter.rs

@@ -1,24 +0,0 @@
-use std::sync::{Arc, RwLock};
-use once_cell::sync::Lazy;
-
-static COUNTER : Lazy<Arc<RwLock<u64>>> = Lazy::new(|| Arc::new(RwLock::new(0)));
-
-#[tauri::command]
-pub fn test_counter() -> u64 {
-  let value : u64;
-  {
-    match COUNTER.write() {
-      Ok(mut counter) => {
-        *counter += 1;
-        value = *counter;
-      },
-      Err(_e) => {
-        value = 0;
-      }
-    }
-  }
-
-  println!("I was invoked from JavaScript! {:?}", value);
-
-  value
-}

crates/desktop/src-tauri/src/endpoints/text_to_image_endpoint.rs

@@ -1,21 +1,21 @@
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
+#[cfg(not(target_os = "macos"))]
+use ml_models::ml::flux::flux_pipeline::{flux_pipeline, FluxModel, FluxPipelineArgs};
+
 use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
-use crate::utils::image::decode_base64_image::decode_base64_image;
-use crate::utils::image::encode_rgb_image_base64_png::encode_rgb_image_base64_png;
+use crate::state::app_dir::{AppDataRoot, AppWeightsDir};
+use crate::stubs::model_cache::ModelCache;
+use crate::stubs::prompt_cache::PromptCache;
+use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
 use image::imageops::FilterType;
 use image::{DynamicImage, ImageReader, RgbImage};
 use log::{error, info};
+
+use ml_weights_registry::weights_registry::flux_weights::{FLUX_DEV, FLUX_SCHNELL, FLUX_SCHNELL_AUTOENCODER, GOOGLE_T5_V1_1_XXL_CONFIG, GOOGLE_T5_V1_1_XXL_MODEL, LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF, LMZ_T5_TOKENIZER_JSON, OPENAI_CLIP_VIT_P14_MODEL, OPENAI_CLIP_VIT_P14_TOKENIZER_JSON};
 use once_cell::sync::Lazy;
 use std::io::Cursor;
-use std::path::PathBuf;
 use tauri::{AppHandle, State};
 
-const RANDOM_SEED: u32 = 42;
-
-const STRENGTH: f64 = 1.0;
+const STRENGTH: f64 = 100.0;
 const PNG_BYTES : &[u8] = include_bytes!("../../binary_includes/1024.png");
 
 // NB: We're just hacking a temporary "text to image" test endpoint leveraging the existing image to image modality.
@@ -33,33 +33,36 @@ pub async fn text_to_image(
   model_config: State<'_, AppConfig>,
   model_cache: State<'_, ModelCache>,
   prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
   app: AppHandle,
+  app_data_root: State<'_, AppDataRoot>,
 ) -> Result<String, String> {
   info!("text_to_image endpoint called.");
 
-  let image = IMAGE.resize(512, 512, FilterType::CatmullRom);
+  let mut image = IMAGE.resize(512, 512, FilterType::CatmullRom);
 
-  let result = do_infer_image(
-    &prompt, 
-    image, 
-    Some(STRENGTH), 
-    &model_config, 
-    &model_cache, 
-    prompt_cache, 
-    app, 
-    &app_data_root
-  ).await;
-  
-  let image = match result {
-    Ok(image) => image,
-    Err(err) => {
-      error!("There was an error generating the image: {:?}", err);
-      return Err(format!("There was an error generating the image: {}", err));
-    }
-  };
+  #[cfg(not(target_os = "macos"))]
+  {
+    let result = text_to_image_impl(
+      &prompt,
+      image,
+      Some(STRENGTH),
+      &model_config,
+      &model_cache,
+      prompt_cache,
+      app_data_root.weights_dir(),
+      app,
+    ).await;
 
-  let bytes = encode_rgb_image_base64_png(image)
+    image = match result {
+      Ok(img) => DynamicImage::ImageRgb8(img),
+      Err(err) => {
+        error!("There was an error generating the image: {:?}", err);
+        return Err(format!("There was an error generating the image: {}", err));
+      }
+    };
+  }
+
+  let bytes = encode_dynamic_image_base64_png(image)
     .map_err(|err| format!("failure to encode image: {:?}", err))?;
 
   info!("Inference successful; image converted to base64, serving back to browser...");
@@ -67,32 +70,44 @@ pub async fn text_to_image(
   Ok(bytes)
 }
 
-async fn do_infer_image(
+#[cfg(not(target_os = "macos"))]
+async fn text_to_image_impl(
   prompt: &str,
   image: DynamicImage,
   strength: Option<f64>,
   config: &AppConfig,
   model_cache: &ModelCache,
   prompt_cache: State<'_, PromptCache>,
+  weights_dir: &AppWeightsDir,
   app: AppHandle,
-  app_data_root: &AppDataRoot,
 ) -> Result<RgbImage, String> {
-
-  let args = Args {
+  
+  let args = FluxPipelineArgs {
     image: &image,
     prompt: prompt.to_string(),
-    uncond_prompt: "".to_string(),
     model_cache,
-    configs: config,
     prompt_cache: &prompt_cache,
-    i2i_strength: strength,
-    cfg_scale: config.cfg_scale,
     app: &app,
-    app_data_root,
-    use_flash_attn: true,
+    flux_model: FluxModel::Schnell,
+    active_blocks: None,
+    use_cuda_stream: Some(true),
+    prefetch_next_batch: Some(false),
+    use_full_gpu: Some(true),
+    img2img_strength: strength,
+    seed: None,
+    use_quantized_model: Some(true),
+    model_config: &config.model_config,
+    flux_dev_or_schnell_path: &weights_dir.weight_path(&FLUX_SCHNELL),
+    flux_schnell_quantized_gguf: &weights_dir.weight_path(&LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF),
+    flux_autoencoder_path: &weights_dir.weight_path(&FLUX_SCHNELL_AUTOENCODER),
+    google_t5_model_path: &weights_dir.weight_path(&GOOGLE_T5_V1_1_XXL_MODEL),
+    google_t5_config_path: &weights_dir.weight_path(&GOOGLE_T5_V1_1_XXL_CONFIG),
+    lmz_t5_tokenizer_path: &weights_dir.weight_path(&LMZ_T5_TOKENIZER_JSON),
+    openai_clip_model_path: &weights_dir.weight_path(&OPENAI_CLIP_VIT_P14_MODEL),
+    openai_clip_tokenizer_json_path: &weights_dir.weight_path(&OPENAI_CLIP_VIT_P14_TOKENIZER_JSON),
   };
 
-  let image = lcm_pipeline(args)
+  let image = flux_pipeline(args)
     .map_err(|err| format!("failure to encode image: {:?}", err))?;
 
   Ok(image)

crates/desktop/src-tauri/src/events/notification_event.rs

@@ -1,6 +1,5 @@
 use serde::Serialize;
 
-
 #[derive(Clone, Copy, Serialize)]
 #[serde(rename_all = "snake_case")]
 pub enum NotificationModelType {
@@ -9,6 +8,9 @@ pub enum NotificationModelType {
   TextTokenizer,
   Unet,
   Vae,
+  Json,
+  ClipEncoder,
+  FluxModel,
 }
 
 #[derive(Clone, Serialize)]
@@ -27,5 +29,17 @@ pub enum NotificationEvent<'a> {
   ModelDownloadComplete {
     model_name: &'a str,
     model_type: NotificationModelType,
-  }
+  },
+  ModelLoading {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
+  ModelInference {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
+  ModelInferenceComplete {
+    model_name: String,
+    model_type: NotificationModelType,
+  },
 }

crates/desktop/src-tauri/src/lib.rs

@@ -1,21 +1,23 @@
 pub mod endpoints;
 pub mod events;
-pub mod ml;
 pub mod state;
+pub mod stubs;
 pub mod threads;
 pub mod transfer;
 pub mod utils;
 
 use crate::endpoints::download_models::download_models;
+use crate::endpoints::flip_image::flip_image;
+use crate::endpoints::inpaint_image_endpoint::inpaint_image;
 use crate::endpoints::realtime_image_endpoint::infer_image;
 use crate::endpoints::remove_background_endpoint::remove_background;
-use crate::endpoints::test_counter::test_counter;
+use crate::endpoints::save_image_endpoint::save_image;
 use crate::endpoints::text_to_image_endpoint::text_to_image;
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
 use crate::state::app_config::AppConfig;
+use crate::stubs::model_cache::ModelCache;
+use crate::stubs::prompt_cache::PromptCache;
 use crate::threads::downloader_thread::downloader_thread;
-use crate::utils::log_environment_details::log_environment_details;
+
 use tauri_plugin_log::Target;
 use tauri_plugin_log::TargetKind;
 
@@ -43,7 +45,10 @@ pub fn run() {
   tauri::Builder::default()
     .plugin(tauri_plugin_log::Builder::new()
       .level(log::LevelFilter::Info)
-      .targets(vec![Target::new(TargetKind::Stdout)])
+      .targets(vec![
+        Target::new(TargetKind::Stdout),
+        Target::new(TargetKind::LogDir { file_name: Some(app_data_root.log_file_name_str().to_string()) }),
+      ])
       .build())
     .setup(|app| {
       // TODO(bt): This is broken on windows
@@ -68,9 +73,12 @@ pub fn run() {
     .manage(app_data_root)
     .invoke_handler(tauri::generate_handler![
       download_models,
+      flip_image,
       infer_image,
+      inpaint_image,
       remove_background,
-      test_counter,
+      save_image,
+      text_to_image,
       text_to_image,
     ])
     .run(tauri::generate_context!())

crates/desktop/src-tauri/src/ml/background_removal/mod.rs

@@ -1 +0,0 @@
-pub mod onnx_session;

crates/desktop/src-tauri/src/state/app_config.rs

@@ -1,32 +1,14 @@
-use crate::ml::model_file::StableDiffusionVersion;
 use crate::state::app_dir::AppDataRoot;
 use crate::state::os_platform::OsPlatform;
 use crate::state::yaml_config::YamlConfig;
-use candle_core::{DType, Device};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use filesys::recursively_find_file_by_name::recursively_find_file_by_name;
-use hf_hub::api::sync::Api;
-use log::info;
-use std::path::PathBuf;
+use crate::stubs::model_config::ModelConfig;
 
 const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
 const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
 
-const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
-const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::V1_5;
 
 /// A central place to configure app-wide details.
 pub struct AppConfig {
-  /// We only support one device type currently, and we can't multiplex 
-  /// over several devices yet.
-  pub device: Device,
-  
-  /// The core numeric type to use for models.
-  pub dtype: DType,
-  
-  pub sd_version: StableDiffusionVersion,
-  pub sd_config: StableDiffusionConfig,
-
   pub image_height: usize,
   pub image_width: usize,
   pub scheduler_steps: usize,
@@ -36,11 +18,10 @@ pub struct AppConfig {
 
   pub cfg_scale: Option<f64>,
   
-  /// Probably shouldn't be here.
-  pub hf_api: Api,
-  
   /// Root location for application data.
   pub app_data_root: AppDataRoot,
+  
+  pub model_config: ModelConfig,
 }
 
 impl AppConfig {
@@ -50,30 +31,6 @@ impl AppConfig {
     
     println!("Configs: {:?}", yaml_configs);
     
-    let device = Device::new_cuda(0)
-      .unwrap_or_else(|e| {
-          info!("CUDA not available ({}), falling back to CPU", e);
-          Device::Cpu
-        });
-
-    
-    let sd_config = match SD_VERSION {
-      StableDiffusionVersion::Turbo => {
-        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
-        StableDiffusionConfig::v1_5(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => {
-        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => {
-        StableDiffusionConfig::sdxl(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-    };
-    
-    let hf_api = Api::new()?;
-    
     let os_platform = OsPlatform::get();
 
     let default_data_root = match os_platform {
@@ -92,18 +49,14 @@ impl AppConfig {
     println!("App data root: {:?}", app_data_root.path());
 
     Ok(Self {
-      device,
-      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
-      sd_version: SD_VERSION,
-      sd_config,
       image_height: yaml_configs.image_height.unwrap_or(DEFAULT_SD_IMAGE_HEIGHT),
       image_width: yaml_configs.image_width.unwrap_or(DEFAULT_SD_IMAGE_WIDTH),
       scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(4),
       scheduler_samples: yaml_configs.scheduler_samples.unwrap_or(15),
       seed: yaml_configs.seed,
       cfg_scale: yaml_configs.cfg_scale,
-      hf_api,
       app_data_root,
+      model_config: ModelConfig::init()?,
     })
   }
 }

crates/desktop/src-tauri/src/state/app_dir.rs

@@ -1,9 +1,10 @@
-use crate::ml::weights_registry::weight_descriptor::WeightDescriptor;
 use crate::state::expanduser::expanduser;
 use crate::state::os_platform::OsPlatform;
 use anyhow::anyhow;
 use directories::UserDirs;
+use ml_weights_registry::weights_registry::weight_descriptor::WeightDescriptor;
 use std::path::{Path, PathBuf};
+use chrono::{DateTime, Local};
 use tempdir::TempDir;
 use tempfile::{Builder, NamedTempFile};
 
@@ -13,6 +14,9 @@ const WEIGHTS_SUBDIRECTORY : &str = "weights";
 
 const TEMPORARY_SUBDIRECTORY : &str = "temp";
 
+/// Note: Tauri appends ".log" to the end of the filename.
+const LOG_FILE_NAME : &str = "application_debug";
+
 /// The path to the application data directory, which includes "asset" and "weights" data.
 #[derive(Clone)]
 pub struct AppDataRoot {
@@ -20,6 +24,8 @@ pub struct AppDataRoot {
   assets_dir: AppAssetsDir,
   weights_dir: AppWeightsDir,
   temp_dir: TemporaryDir,
+  log_file_name: PathBuf,
+  log_file_name_string: String,
 }
 
 #[derive(Clone)]
@@ -70,12 +76,19 @@ impl AppDataRoot {
     let assets_dir = AppAssetsDir::create_existing(dir.join(ASSETS_SUBDIRECTORY))?;
     let weights_dir = AppWeightsDir::create_existing(dir.join(WEIGHTS_SUBDIRECTORY))?;
     let temp_dir = TemporaryDir::create_existing(dir.join(TEMPORARY_SUBDIRECTORY))?;
+    let log_file_name = dir.join(LOG_FILE_NAME);
+    let log_file_name_string = log_file_name
+        .to_str()
+        .ok_or(anyhow!("couldn't convert log path to str"))?
+        .to_string();
 
     Ok(Self {
       path: dir,
       assets_dir,
       weights_dir,
       temp_dir,
+      log_file_name,
+      log_file_name_string,
     })
   }
   
@@ -94,6 +107,14 @@ impl AppDataRoot {
   pub fn path(&self) -> &Path {
     &self.path
   }
+
+  pub fn log_file_name(&self) -> &Path {
+    &self.log_file_name
+  }
+
+  pub fn log_file_name_str(&self) -> &str {
+    &self.log_file_name_string
+  }
 }
 
 impl AppAssetsDir {
@@ -117,6 +138,18 @@ impl AppAssetsDir {
   pub fn path(&self) -> &Path {
     &self.path
   }
+  
+  /// Return the current date asset directory
+  /// If it doesn't exist, create it.
+  pub fn make_or_get_current_date_dir(&self) -> anyhow::Result<PathBuf> {
+    let format: DateTime<Local> = Local::now();
+    let date_directory_name = format.format("%Y-%d-%m").to_string();
+    let full_path = self.path.join(date_directory_name);
+    if !full_path.exists() {
+      std::fs::create_dir(&full_path)?;
+    }
+    Ok(full_path)
+  }
 }
 
 impl AppWeightsDir {

crates/desktop/src-tauri/src/stubs/mod.rs

@@ -0,0 +1,3 @@
+pub mod model_cache;
+pub mod model_config;
+pub mod prompt_cache;

crates/desktop/src-tauri/src/stubs/model_cache.rs

@@ -0,0 +1,14 @@
+#[cfg(not(target_os = "macos"))]
+pub use ml_models::ml::model_cache::ModelCache;
+
+#[cfg(not(not(target_os = "macos")))]
+pub struct ModelCache {
+  // Intentionally left blank
+}
+
+#[cfg(not(not(target_os = "macos")))]
+impl ModelCache {
+  pub fn new() -> Self {
+    Self {}
+  }
+}

crates/desktop/src-tauri/src/stubs/model_config.rs

@@ -0,0 +1,14 @@
+#[cfg(not(target_os = "macos"))]
+pub use ml_models::ml::model_config::ModelConfig;
+
+#[cfg(not(not(target_os = "macos")))]
+pub struct ModelConfig {
+  // Intentionally left blank
+}
+
+#[cfg(not(not(target_os = "macos")))]
+impl ModelConfig {
+  pub fn init() -> anyhow::Result<Self> {
+    Ok(Self {})
+  }
+}

crates/desktop/src-tauri/src/stubs/prompt_cache.rs

@@ -0,0 +1,15 @@
+
+#[cfg(not(target_os = "macos"))]
+pub use ml_models::ml::prompt_cache::PromptCache;
+
+#[cfg(not(not(target_os = "macos")))]
+pub struct PromptCache {
+  // Intentionally left blank
+}
+
+#[cfg(not(not(target_os = "macos")))]
+impl PromptCache {
+  pub fn with_capacity(capacity: usize) -> anyhow::Result<Self> {
+    Ok(Self {})
+  }
+}

crates/desktop/src-tauri/src/threads/downloader_thread.rs

@@ -1,9 +1,11 @@
 use crate::events::notification_event::{NotificationEvent, NotificationModelType};
 use crate::state::app_dir::AppDataRoot;
 use crate::transfer::download::{download_async, ProgressUpdate};
-use futures::stream::FuturesUnordered;
 use futures::StreamExt;
+use futures::stream::FuturesUnordered;
 use log::{debug, error, info};
+use ml_weights_registry::weights_registry::weight_descriptor::WeightFunction;
+use ml_weights_registry::weights_registry::weights::{CLIP_JSON, DIS_MEDIUM_ONNX, LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2, SDXL_TURBO_UNET, SDXL_TURBO_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
 use std::collections::VecDeque;
 use std::path::PathBuf;
 use std::sync::mpsc::Receiver;
@@ -11,8 +13,7 @@ use std::sync::{Arc, Mutex};
 use tauri::{AppHandle, Emitter};
 use tempfile::NamedTempFile;
 use tokio::task::JoinHandle;
-use crate::ml::weights_registry::weight_descriptor::WeightFunction;
-use crate::ml::weights_registry::weights::{CLIP_JSON, DIS_MEDIUM_ONNX, LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2, SDXL_TURBO_UNET, SDXL_TURBO_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
+use ml_weights_registry::weights_registry::flux_weights::{FLUX_SCHNELL, FLUX_SCHNELL_AUTOENCODER, GOOGLE_T5_V1_1_XXL_CONFIG, GOOGLE_T5_V1_1_XXL_MODEL, LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF, LMZ_T5_TOKENIZER_JSON, OPENAI_CLIP_VIT_P14_MODEL, OPENAI_CLIP_VIT_P14_TOKENIZER_JSON};
 
 const NOTIFICATION_CHANNEL_NAME : &str = "notification";
 const MAX_FILES : usize = 8;
@@ -50,6 +51,16 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
   //download_queue.push_back(SDXL_TURBO_VAE); // TODO(bt): REMOVE
   //download_queue.push_back(SDXL_TURBO_UNET); // TODO(bt): REMOVE
 
+  // Flux:
+  download_queue.push_back(FLUX_SCHNELL);
+  download_queue.push_back(FLUX_SCHNELL_AUTOENCODER);
+  download_queue.push_back(LMZ_CANDLE_FLUX_SCHNELL_QUANTIZED_GGUF);
+  download_queue.push_back(GOOGLE_T5_V1_1_XXL_MODEL);
+  download_queue.push_back(GOOGLE_T5_V1_1_XXL_CONFIG);
+  download_queue.push_back(LMZ_T5_TOKENIZER_JSON);
+  download_queue.push_back(OPENAI_CLIP_VIT_P14_MODEL);
+  download_queue.push_back(OPENAI_CLIP_VIT_P14_TOKENIZER_JSON);
+
   let mut handles = FuturesUnordered::new();
 
   while let Some(model) = download_queue.pop_front() {
@@ -82,6 +93,7 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
       WeightFunction::Unet => NotificationModelType::Unet,
       WeightFunction::Vae => NotificationModelType::Vae,
       WeightFunction::ImageSegmentation => NotificationModelType::ImageSegmentation,
+      WeightFunction::ImageGeneration => NotificationModelType::ImageSegmentation, // TODO FIX
     };
 
     let result = app.emit(NOTIFICATION_CHANNEL_NAME, NotificationEvent::ModelDownloadStarted {

crates/desktop/src-tauri/src/utils/image/encode_rgb_image_base64_png.rs

@@ -1,7 +1,6 @@
-use crate::ml::image::tensor_to_image_buffer::RgbImage;
 use base64::prelude::BASE64_STANDARD;
 use base64::Engine;
-use image::ImageFormat;
+use image::{ImageFormat, RgbImage};
 use std::io::Cursor;
 
 pub fn encode_rgb_image_base64_png(

crates/desktop/src-tauri/src/utils/log_environment_details.rs

@@ -1,11 +1,11 @@
 use log::info;
-use nvidia_checker::{get_cuda_version, get_cudnn_version, get_kernel_version, get_nvidia_driver_version, get_os_version, get_tensorrt_version};
+//use nvidia_checker::{get_cuda_version, get_cudnn_version, get_kernel_version, get_nvidia_driver_version, get_os_version, get_tensorrt_version};
 
 pub fn log_environment_details() {
-  info!("NVIDIA driver version: {}", get_nvidia_driver_version(false));
-  info!("CUDA version: {}", get_cuda_version(false));
-  info!("cuDNN version: {}", get_cudnn_version(false));
-  info!("TensorRT version: {}", get_tensorrt_version(false));
-  info!("OS version: {}", get_os_version(false));
-  info!("Kernel version: {}", get_kernel_version(false));
+  //info!("NVIDIA driver version: {}", get_nvidia_driver_version(false));
+  //info!("CUDA version: {}", get_cuda_version(false));
+  //info!("cuDNN version: {}", get_cudnn_version(false));
+  //info!("TensorRT version: {}", get_tensorrt_version(false));
+  //info!("OS version: {}", get_os_version(false));
+  //info!("Kernel version: {}", get_kernel_version(false));
 }

crates/desktop/src-tauri/tauri.conf.json

@@ -32,9 +32,6 @@
       "icons/128x128@2x.png",
       "icons/icon.icns",
       "icons/icon.ico"
-    ],
-    "resources": {
-      "../../../windows_dlls/*.dll": "."
-    }
+    ]
   }
 }

crates/lib/actix_cors_configs/src/configs/artcraft_2d.rs

@@ -0,0 +1,13 @@
+use actix_cors::Cors;
+use crate::util::netlify_branch_domain_matches::netlify_branch_domain_matches;
+
+pub fn add_artcraft_2d(cors: Cors, _is_production: bool) -> Cors {
+  cors
+      // Hypothetical domains
+      .allowed_origin("https://2d.storyteller.ai")
+      .allowed_origin("https://2d.getartcraft.com")
+      // Netlify project
+      .allowed_origin_fn(|origin, _req_head| {
+        netlify_branch_domain_matches(origin, "storyteller-2d.netlify.app")
+      })
+}

crates/lib/actix_cors_configs/src/configs/artcraft_3d.rs

@@ -0,0 +1,13 @@
+use actix_cors::Cors;
+use crate::util::netlify_branch_domain_matches::netlify_branch_domain_matches;
+
+pub fn add_artcraft_3d(cors: Cors, _is_production: bool) -> Cors {
+  cors
+      // Hypothetical domains
+      .allowed_origin("https://3d.storyteller.ai")
+      .allowed_origin("https://3d.getartcraft.com")
+      // Netlify project
+      .allowed_origin_fn(|origin, _req_head| {
+        netlify_branch_domain_matches(origin, "storyteller-3d.netlify.app")
+      })
+}

crates/lib/actix_cors_configs/src/configs/mod.rs

@@ -1,3 +1,5 @@
+pub (crate) mod artcraft_2d;
+pub (crate) mod artcraft_3d;
 pub (crate) mod development_only;
 pub (crate) mod fakeyou;
 pub (crate) mod legacy;

crates/lib/actix_cors_configs/src/cors.rs

@@ -2,7 +2,8 @@ use actix_cors::Cors;
 use log::info;
 
 use reusable_types::server_environment::ServerEnvironment;
-
+use crate::configs::artcraft_2d::add_artcraft_2d;
+use crate::configs::artcraft_3d::add_artcraft_3d;
 use crate::configs::development_only::add_development_only;
 use crate::configs::fakeyou::{add_fakeyou, add_fakeyou_dev_proxy};
 use crate::configs::legacy::{add_legacy_storyteller_stream, add_legacy_trumped, add_legacy_vocodes, add_power_stream};
@@ -40,6 +41,10 @@ fn do_build_cors_config(is_production: bool) -> Cors {
   cors = add_storyteller_board(cors, is_production);
   cors = add_storyteller_render(cors, is_production);
 
+  // Artcraft
+  cors = add_artcraft_2d(cors, is_production);
+  cors = add_artcraft_3d(cors, is_production);
+
   // Legacy
   cors = add_legacy_trumped(cors, is_production);
   cors = add_power_stream(cors, is_production);

crates/lib/api_clients/cloud_storage/src/bucket_client.rs

@@ -234,17 +234,31 @@ impl BucketClient {
       .map(|s| s.to_string())
       .ok_or(anyhow!("could not convert object path to string"))?;
 
-    info!("downloading from bucket: {:?}", &object_path_str);
+    info!("creating file for bucket download: {:?}", filesystem_path.as_ref());
 
-    let mut output_file = tokio::fs::File::create(filesystem_path).await?;
+    let mut output_file = File::create(filesystem_path).await?;
 
-    let status_code = self.bucket.get_object_to_writer(&object_path_str, &mut output_file).await?;
+    let result = self.bucket.get_object_to_writer(&object_path_str, &mut output_file).await;
+
+    info!("downloading from bucket (named '{}'), path: {}", &self.bucket.name, &object_path_str);
+
+    let status_code = match result {
+      Ok(status_code) => status_code,
+      Err(err) => {
+        return bail!("Error downloading from bucket (named '{}'): {:?}", &self.bucket.name, err)
+      }
+    };
 
     match status_code {
-      404 => bail!("File not found in bucket: {}", &object_path_str),
-      _ => {},
+      404 => {
+        error!("File not found in bucket (named '{}'), path: {}", &self.bucket.name, &object_path_str);
+        bail!("File not found in bucket: {}", &object_path_str)
+      },
+      _ => {
+        info!("download code: {}", status_code);
+      },
     }
-    info!("download code: {}", status_code);
+
     Ok(())
   }
 }

crates/lib/api_clients/openai_sora_client/Cargo.toml

@@ -0,0 +1,35 @@
+[package]
+name = "openai_sora_client"
+edition = "2021"
+version = "0.0.1"
+publish = false
+
+[package.metadata.cargo-udeps.ignore]
+normal = ["my-workspace-hack"]
+
+[lib]
+name = "openai_sora_client"
+path = "src/lib.rs"
+
+[dependencies]
+
+# Workspace hack (faster compile times)
+my-workspace-hack = { workspace =  true }
+
+# Internal
+crockford = { path = "../../data/crockford" }
+errors = { path = "../../errors" }
+subprocess_common = { path = "../../subprocess_common" }
+
+# External
+anyhow = { workspace = true }
+log = { workspace = true }
+reqwest = { workspace = true, features = ["multipart"] }
+serde = { workspace = true }
+serde_derive = { workspace = true }
+serde_json = { workspace = true }
+tokio = { version = "1.38.0", features = ["macros"] }
+
+[dev-dependencies]
+testing = { path = "../../testing" }
+filesys = { path = "../../files/filesys" }

crates/lib/api_clients/openai_sora_client/src/credentials.rs

@@ -0,0 +1,43 @@
+use reqwest::RequestBuilder;
+use errors::AnyhowResult;
+// NB: It appears that the sentinel may require a matching user agent.
+pub (crate) const USER_AGENT : &str = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/127.0.0.0 Safari/537.36";
+
+#[derive(Clone)]
+pub struct SoraCredentials {
+  pub bearer_token: String,
+  pub cookie: String,
+  pub sentinel: Option<String>,
+}
+
+impl SoraCredentials {
+  pub fn authorization_header_value(&self) -> String {
+    match self.bearer_token.get(0..6) {
+      Some("bearer") | Some("Bearer") => self.bearer_token.clone(),
+      _ => "Bearer ".to_owned() + &self.bearer_token,
+    }
+  }
+
+  pub fn from_env() -> AnyhowResult<Self> {
+    let bearer = std::env::var("SORA_BEARER_TOKEN")?;
+    let cookie = std::env::var("SORA_COOKIE")?;
+    let sentinel = std::env::var("SORA_SENTINEL").ok();
+
+    Ok(Self { bearer_token: bearer, cookie, sentinel })
+  }
+
+  pub fn add_credential_headers_to_request(&self, request: RequestBuilder) -> RequestBuilder {
+    let bearer_header = self.authorization_header_value();
+
+    let mut request = request
+        .header("User-Agent", USER_AGENT)
+        .header("Cookie", &self.cookie)
+        .header("Authorization", bearer_header);
+
+    if let Some(sentinel) = &self.sentinel {
+      request = request.header("OpenAI-Sentinel-Token", sentinel);
+    }
+
+    request
+  }
+}

crates/lib/api_clients/openai_sora_client/src/image_gen/common.rs

@@ -0,0 +1,48 @@
+
+#[derive(Debug, Clone)]
+pub struct SoraImageGenResponse {
+  pub task_id: String,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum NumImages {
+  One,
+  Two,
+  Four
+}
+
+impl NumImages {
+  pub fn as_count(&self) -> usize {
+    match self {
+      NumImages::One => 1,
+      NumImages::Two => 2,
+      NumImages::Four => 4,
+    }
+  }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ImageSize {
+  Square,
+  Wide,
+  Tall,
+}
+
+impl ImageSize {
+  // TODO: Verify these dimensions. 480x480 is correct.
+  pub fn as_dimensions(&self) -> (u16, u16) {
+    match self {
+      ImageSize::Square => (480, 480),
+      ImageSize::Wide => (854, 480),
+      ImageSize::Tall => (480, 854),
+    }
+  }
+
+  pub fn as_width(&self) -> u16 {
+    self.as_dimensions().0
+  }
+
+  pub fn as_height(&self) -> u16 {
+    self.as_dimensions().1
+  }
+}

crates/lib/api_clients/openai_sora_client/src/image_gen/image_gen_http_request.rs

@@ -0,0 +1,130 @@
+use anyhow::anyhow;
+use serde_derive::{Deserialize, Serialize};
+use errors::AnyhowResult;
+use crate::credentials::SoraCredentials;
+
+const SORA_IMAGE_GEN_URL: &str = "https://sora.com/backend/video_gen";
+
+/// This user agent is tied to the sentinel generation. If we need to change it, we may need to change sentinel generation too.
+const USER_AGENT : &str = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/127.0.0.0 Safari/537.36";
+
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) enum VideoGenType {
+  ImageGen,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) enum OperationType {
+  /// Simple prompt without reference images
+  SimpleCompose,
+  /// Prompt with reference images
+  Remix,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) enum InpaintItemType {
+  Image,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) struct InpaintItem {
+  pub r#type: InpaintItemType,
+  /// Typically "0"
+  pub frame_index: u32,
+  /// Unknown field; defaults to null
+  pub preset_id: Option<String>,
+  /// Unknown field; defaults to null
+  pub generation_id: Option<String>,
+  /// Token identifier of the item, eg. "media_01jqt9vt20erx9zvryf3v1pecx"
+  pub upload_media_id: String,
+  /// Typically "0"
+  pub source_start_frame: u32,
+  /// Typically "0"
+  pub source_end_frame: u32,
+  /// Unknown field; defaults to null
+  pub crop_bounds: Option<String>,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) struct RawSoraImageGenRequest {
+  /// eg. "image_gen"
+  pub r#type: VideoGenType,
+  /// eg. "simple_compose"
+  pub operation: OperationType,
+  /// The user's raw prompt
+  pub prompt: String,
+  /// The number of variants to generate, eg. 2
+  pub n_variants: usize,
+  /// The width of the image, eg. 480
+  pub width: u16,
+  /// The height of the image, eg. 480
+  pub height: u16,
+  /// The number of frames to generate, eg. 1
+  pub n_frames: u8,
+  /// The items to inpaint, eg. []
+  pub inpaint_items: Vec<InpaintItem>,
+}
+
+#[derive(Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) struct RawSoraResponse {
+  /// eg. "task_01jqsz9dsae9tvjygf1abrv3xf"
+  pub id: String,
+  /// not known
+  pub priority: Option<String>,
+}
+
+// https://sora.com/backend/notif?limit=100&before=task_01jqpg8qghenet0rw2a79p0vbn
+// error
+// message	"Missing bearer authentication in header"
+// type	"invalid_request_error"
+// param	null
+// code	null
+#[derive(Serialize)]
+#[serde(rename_all = "snake_case")]
+pub (crate) struct RawSoraErrorResponse {
+  error: String,
+  message: String,
+  r#type: String,
+  param: Option<String>,
+  code: Option<String>,
+}
+
+/// Don't expose the internal request implementation as there are only a few "correct" ways to call the API.
+pub (crate) async fn image_gen_http_request(sora_request: RawSoraImageGenRequest, credentials: &SoraCredentials) -> AnyhowResult<RawSoraResponse> {
+  let client = reqwest::Client::new();
+
+  let sentinel = match credentials.sentinel.as_deref() {
+    None => return Err(anyhow!("Calls to image gen require a sentinel in the credentials payload.")),
+    Some(sentinel) => sentinel,
+  };
+
+  let http_request = client.post(SORA_IMAGE_GEN_URL)
+      .header("User-Agent", USER_AGENT)
+      .header("Cookie", &credentials.cookie)
+      .header("Authorization", credentials.authorization_header_value())
+      .header("Content-Type", "application/json")
+      .header("OpenAI-Sentinel-Token", sentinel);
+
+  let http_request = http_request.json(&sora_request).build()?;
+
+  let response = client.execute(http_request)
+      .await?;
+
+  let status = response.status();
+
+  let response_body = &response.text().await?;
+
+  if status != reqwest::StatusCode::OK {
+    return Err(anyhow::anyhow!("the request failed; status = {:?}, message = {:?}", status, response_body));
+  }
+
+  let response = serde_json::from_str(response_body)?;
+
+  Ok(response)
+}

crates/lib/api_clients/openai_sora_client/src/image_gen/mod.rs

@@ -0,0 +1,4 @@
+pub (crate) mod image_gen_http_request;
+pub mod common;
+pub mod sora_image_gen_remix;
+pub mod sora_image_gen_simple;

crates/lib/api_clients/openai_sora_client/src/image_gen/sora_image_gen_remix.rs

@@ -0,0 +1,87 @@
+use errors::AnyhowResult;
+use crate::credentials::SoraCredentials;
+use crate::image_gen::common::{ImageSize, NumImages, SoraImageGenResponse};
+use crate::image_gen::image_gen_http_request::{image_gen_http_request, InpaintItem, InpaintItemType, OperationType, RawSoraImageGenRequest, VideoGenType};
+
+pub struct SoraImageGenRemixRequest<'a> {
+  pub prompt: String,
+  pub num_images: NumImages,
+  pub image_size: ImageSize,
+  pub sora_media_tokens: Vec<String>,
+  pub credentials: &'a SoraCredentials,
+}
+
+/// The "remix" commands let you supply additional images as context.
+/// Sora "media tokens" of previously uploaded images must be supplied.
+pub async fn sora_image_gen_remix(request: SoraImageGenRemixRequest<'_>) -> AnyhowResult<SoraImageGenResponse> {
+  let args = RawSoraImageGenRequest {
+    r#type: VideoGenType::ImageGen,
+    operation: OperationType::Remix,
+    prompt: request.prompt,
+    n_variants: request.num_images.as_count(),
+    width: request.image_size.as_width(),
+    height: request.image_size.as_height(),
+    n_frames: 1,
+    inpaint_items: request.sora_media_tokens.into_iter().map(|token| {
+      InpaintItem {
+        r#type: InpaintItemType::Image,
+        frame_index: 0,
+        preset_id: None,
+        generation_id: None,
+        upload_media_id: token,
+        source_start_frame: 0,
+        source_end_frame: 0,
+        crop_bounds: None,
+      }
+    }).collect(),
+  };
+
+  // TODO: Error handling.
+  let result = image_gen_http_request(args, &request.credentials).await?;
+
+  Ok(SoraImageGenResponse {
+    task_id: result.id,
+  })
+}
+
+#[cfg(test)]
+mod tests {
+  use std::fs::read_to_string;
+  use errors::AnyhowResult;
+  use testing::test_file_path::test_file_path;
+  use crate::credentials::SoraCredentials;
+  use crate::image_gen::common::{ImageSize, NumImages};
+  use crate::image_gen::sora_image_gen_remix::{sora_image_gen_remix, SoraImageGenRemixRequest};
+
+  #[ignore] // You can manually run "ignore" tests in the IDE, but they won't run in CI.
+  #[tokio::test]
+  pub async fn manual_test() -> AnyhowResult<()> {
+    let sentinel = read_to_string(test_file_path("test_data/temp/sentinel.txt")?)?;
+    let sentinel = sentinel.trim().to_string();
+
+    let cookie = read_to_string(test_file_path("test_data/temp/cookie.txt")?)?;
+    let cookie = cookie.trim().to_string();
+
+    let bearer = read_to_string(test_file_path("test_data/temp/bearer.txt")?)?;
+    let bearer = bearer.trim().to_string();
+
+    let creds = SoraCredentials {
+      bearer_token: bearer,
+      cookie,
+      sentinel: Some(sentinel),
+    };
+
+    let response = sora_image_gen_remix(SoraImageGenRemixRequest {
+      prompt: "Match the pose and scene layout from the uploaded image exactly. A smart dog getting kisses from another dog. Brick building. Anime style".to_string(),
+      num_images: NumImages::One,
+      image_size: ImageSize::Square,
+      sora_media_tokens: vec!["media_01jqyhrz4detyvtzwp2p4j63ad".to_string()],
+      credentials: &creds,
+    }).await?;
+
+    println!("task_id: {}", response.task_id);
+
+    assert!(response.task_id.starts_with("task_"));
+    Ok(())
+  }
+}

crates/lib/api_clients/openai_sora_client/src/image_gen/sora_image_gen_simple.rs

@@ -0,0 +1,72 @@
+use errors::AnyhowResult;
+use crate::credentials::SoraCredentials;
+use crate::image_gen::common::{ImageSize, NumImages, SoraImageGenResponse};
+use crate::image_gen::image_gen_http_request::{image_gen_http_request, OperationType, RawSoraImageGenRequest, VideoGenType};
+
+pub struct SoraImageGenSimpleRequest<'a> {
+  pub prompt: String,
+  pub num_images: NumImages,
+  pub image_size: ImageSize,
+  pub credentials: &'a SoraCredentials,
+}
+
+pub async fn sora_image_gen_simple(args: SoraImageGenSimpleRequest<'_>) -> AnyhowResult<SoraImageGenResponse> {
+  let sora_request = RawSoraImageGenRequest {
+    r#type: VideoGenType::ImageGen,
+    operation: OperationType::SimpleCompose,
+    prompt: args.prompt,
+    n_variants: args.num_images.as_count(),
+    width: args.image_size.as_width(),
+    height: args.image_size.as_height(),
+    n_frames: 1,
+    inpaint_items: vec![],
+  };
+
+  // TODO: Error handling.
+  let result = image_gen_http_request(sora_request, args.credentials).await?;
+
+  Ok(SoraImageGenResponse {
+    task_id: result.id,
+  })
+}
+
+#[cfg(test)]
+mod tests {
+  use std::fs::read_to_string;
+  use errors::AnyhowResult;
+  use testing::test_file_path::test_file_path;
+  use crate::credentials::SoraCredentials;
+  use crate::image_gen::common::{ImageSize, NumImages};
+  use crate::image_gen::sora_image_gen_simple::{sora_image_gen_simple, SoraImageGenSimpleRequest};
+
+  #[ignore] // You can manually run "ignore" tests in the IDE, but they won't run in CI.
+  #[tokio::test]
+  pub async fn manual_test() -> AnyhowResult<()> {
+    let sentinel = read_to_string(test_file_path("test_data/temp/sentinel.txt")?)?;
+    let sentinel = sentinel.trim().to_string();
+
+    let cookie = read_to_string(test_file_path("test_data/temp/cookie.txt")?)?;
+    let cookie = cookie.trim().to_string();
+
+    let bearer = read_to_string(test_file_path("test_data/temp/bearer.txt")?)?;
+    let bearer = bearer.trim().to_string();
+
+    let creds = SoraCredentials {
+      bearer_token: bearer,
+      cookie,
+      sentinel: Some(sentinel),
+    };
+
+    let response = sora_image_gen_simple(SoraImageGenSimpleRequest {
+      prompt: "A pirate and a ninja fight in a battle inside a UFO. Fully photo realistic, lifelike, lens flare".to_string(),
+      num_images: NumImages::One,
+      image_size: ImageSize::Square,
+      credentials: &creds,
+    }).await?;
+
+    println!("task_id: {}", response.task_id);
+
+    assert!(response.task_id.starts_with("task_"));
+    Ok(())
+  }
+}

crates/lib/api_clients/openai_sora_client/src/job/mod.rs

@@ -0,0 +1 @@
+pub mod sora_job_status;

crates/lib/api_clients/openai_sora_client/src/job/sora_job_status.rs

@@ -0,0 +1,64 @@
+const USER_AGENT : &str = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/127.0.0.0 Safari/537.36";
+
+// https://sora.com/backend/notif?limit=100&before=task_01jqwwrkvgeqp8jsf5mqk1jceh
+
+use crate::credentials::SoraCredentials;
+
+pub async fn sora_job_status(
+  task_id: &str,
+  credentials: &SoraCredentials,
+) -> anyhow::Result<()> {
+  let url = format!("https://sora.com/backend/notif?limit=100&before={}", task_id);
+
+  let client = reqwest::Client::new();
+
+  let request = client
+    .get(url)
+    .header("Authorization", &credentials.bearer_token)
+    .header("User-Agent", USER_AGENT)
+    .header("Cookie", &credentials.cookie);
+
+  let response = request.send().await?;
+
+  let json_response = &response.text().await?;
+
+  println!(" >>> response = {:?}", json_response);
+
+  //let response = serde_json::from_str(json_response)?;
+
+  Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+  use std::fs::read_to_string;
+  use errors::AnyhowResult;
+  use testing::test_file_path::test_file_path;
+  use crate::credentials::SoraCredentials;
+  use crate::job::sora_job_status::sora_job_status;
+
+  #[ignore]
+  #[tokio::test]
+  pub async fn manual_test() -> AnyhowResult<()> {
+    let sentinel = read_to_string(test_file_path("test_data/temp/sentinel.txt")?)?;
+    let sentinel = sentinel.trim().to_string();
+
+    let cookie = read_to_string(test_file_path("test_data/temp/cookie.txt")?)?;
+    let cookie = cookie.trim().to_string();
+
+    let bearer = read_to_string(test_file_path("test_data/temp/bearer.txt")?)?;
+    let bearer = bearer.trim().to_string();
+
+    let creds = SoraCredentials {
+      bearer_token: bearer,
+      cookie,
+      sentinel: Some(sentinel),
+    };
+
+    let task_id = "task_01jqwwrkvgeqp8jsf5mqk1jceh";
+
+    let result = sora_job_status(task_id, &creds).await?;
+
+    Ok(())
+  }
+}

crates/lib/api_clients/openai_sora_client/src/lib.rs

@@ -0,0 +1,4 @@
+pub mod credentials;
+pub mod image_gen;
+pub mod job;
+pub mod upload;

crates/lib/api_clients/openai_sora_client/src/upload/mod.rs

@@ -0,0 +1,3 @@
+pub (crate) mod upload_media_http_request;
+pub mod upload_media_from_bytes;
+pub mod upload_media_from_file;

crates/lib/api_clients/openai_sora_client/src/upload/upload_media_from_bytes.rs

@@ -0,0 +1,29 @@
+use std::path::{Path, PathBuf};
+use errors::AnyhowResult;
+use crate::credentials::SoraCredentials;
+use crate::upload::upload_media_http_request::{upload_media_http_request, SoraMediaUploadResponse};
+
+pub struct SoraMediaUploadRequest<'a> {
+  pub file_path: String,
+  pub credentials: &'a SoraCredentials,
+}
+
+/// Upload bytes.
+/// The underlying reqwest lib needs to own the bytes, so we can't pass as a reference.
+pub async fn sora_media_upload_from_bytes<P: AsRef<Path>>(bytes: Vec<u8>, file_name: String, creds: &SoraCredentials) -> AnyhowResult<SoraMediaUploadResponse> {
+  let file_path = PathBuf::from(&file_name);
+
+  // TODO: Read file magic bytes first, then fall back to this.
+  let mime_type = match file_path.extension().and_then(|e| e.to_str()) {
+    Some("jpg") | Some("jpeg") => "image/jpeg",
+    Some("png") => "image/png",
+    // Some("webp") => "image/webp",
+    // Some("gif") => "image/gif",
+    // Some("mp4") => "video/mp4",
+    // Some("mov") => "video/quicktime",
+    // Some("webm") => "video/webm",
+    _ => "application/octet-stream",
+  };
+
+  Ok(upload_media_http_request(bytes, file_name, mime_type, creds).await?)
+}

crates/lib/api_clients/openai_sora_client/src/upload/upload_media_from_file.rs

@@ -0,0 +1,81 @@
+use std::path::Path;
+use tokio::fs::File;
+use tokio::io::AsyncReadExt;
+use errors::AnyhowResult;
+use crate::credentials::SoraCredentials;
+use crate::upload::upload_media_http_request::{upload_media_http_request, SoraMediaUploadResponse};
+
+/// Try to prevent buffer reallocations.
+/// There's a better way to implement this.
+const INITIAL_BUFFER_SIZE : usize = 1024*1024;
+
+pub struct SoraMediaUploadRequest<'a> {
+  pub file_path: String,
+  pub credentials: &'a SoraCredentials,
+}
+
+pub async fn sora_media_upload_from_file<P: AsRef<Path>>(file_path: P, creds: &SoraCredentials) -> AnyhowResult<SoraMediaUploadResponse> {
+  let mut file = File::open(&file_path).await?;
+  let mut buffer = Vec::with_capacity(INITIAL_BUFFER_SIZE);
+  file.read_to_end(&mut buffer).await?;
+
+  let filename = file_path.as_ref().file_name()
+    .ok_or_else(|| anyhow::anyhow!("Invalid file path"))?
+    .to_string_lossy()
+    .to_string();
+
+  // TODO: Read file magic bytes first, then fall back to this.
+  let mime_type = match file_path.as_ref().extension().and_then(|e| e.to_str()) {
+    Some("jpg") | Some("jpeg") => "image/jpeg",
+    Some("png") => "image/png",
+    // Some("webp") => "image/webp",
+    // Some("gif") => "image/gif",
+    // Some("mp4") => "video/mp4",
+    // Some("mov") => "video/quicktime",
+    // Some("webm") => "video/webm",
+    _ => "application/octet-stream",
+  };
+
+  Ok(upload_media_http_request(buffer, filename, mime_type, creds).await?)
+}
+
+#[cfg(test)]
+mod tests {
+  use std::fs::read_to_string;
+  use errors::AnyhowResult;
+  use testing::test_file_path::test_file_path;
+  use crate::credentials::SoraCredentials;
+  use crate::upload::upload_media_from_file::sora_media_upload_from_file;
+
+  #[ignore] // You can manually run "ignore" tests in the IDE, but they won't run in CI.
+  #[tokio::test]
+  pub async fn manual_test() -> AnyhowResult<()> {
+    let cookie = read_to_string(test_file_path("test_data/temp/cookie.txt")?)?;
+    let cookie = cookie.trim().to_string();
+
+    let bearer = read_to_string(test_file_path("test_data/temp/bearer.txt")?)?;
+    let bearer = bearer.trim().to_string();
+
+    let image_path = test_file_path("test_data/image/juno.jpg")?; // media_01jqyqgqpwf40tkcapq5bmaz5d
+
+    let creds = SoraCredentials {
+      bearer_token: bearer,
+      cookie,
+      sentinel: None,
+    };
+
+    let response = sora_media_upload_from_file(
+      image_path,
+      &creds,
+    ).await?;
+
+    println!("media: {:?}", response);
+
+    println!("media_id: {}", response.id);
+    println!("media_url: {}", response.url);
+
+    assert!(response.id.starts_with("media_"));
+
+    Ok(())
+  }
+}

crates/lib/api_clients/openai_sora_client/src/upload/upload_media_http_request.rs

@@ -0,0 +1,81 @@
+use crate::credentials::{SoraCredentials, USER_AGENT};
+use errors::AnyhowResult;
+use reqwest::multipart::{Form, Part};
+use reqwest::Client;
+use serde::Deserialize;
+
+const SORA_UPLOAD_MEDIA_URL: &str = "https://sora.com/backend/uploads";
+
+// Content-Disposition: form-data; name="file"; filename="imagename.webp"
+
+/*
+Response:
+{
+  "id": "media_01jqt9vt20erx9zvryf3v1pecx",
+  "type": "image",
+  "created_at": "2025-04-02T03:49:04.185867Z",
+  "filename": "Rith_essa.webp",
+  "extension": "jpeg",
+  "mime_type": "image/jpeg",
+  "url": "https://videos.openai.com/vg-assets/assets%2Fclient_upload%2Fmedia%2Fuser-9wf6JFdgRSJLjvSJ53LNAbV8%2Fmedia_01jqt9vt20erx9zvryf3v1pecx.jpeg?st=2025-04-02T02%3A39%3A01Z&se=2025-04-08T03%3A39%3A01Z&sks=b&skt=2025-04-02T02%3A39%3A01Z&ske=2025-04-08T03%3A39%3A01Z&sktid=a48cca56-e6da-484e-a814-9c849652bcb3&skoid=3d249c53-07fa-4ba4-9b65-0bf8eb4ea46a&skv=2019-02-02&sv=2018-11-09&sr=b&sp=r&spr=https%2Chttp&sig=%2BdnTQyw9OFF2kaVoEReT2cVfTqmKjGt18m7P4wKCWCI%3D&az=oaivgprodscus",
+  "width": 1200,
+  "height": 675,
+  "duration_sec": null,
+  "n_frames": 1,
+  "size_bytes": 118501,
+  "thumbnail_url": "https://videos.openai.com/vg-assets/assets%2Fclient_upload%2Fmedia%2Fuser-9wf6JFdgRSJLjvSJ53LNAbV8%2Fmedia_01jqt9vt20erx9zvryf3v1pecx.jpg?st=2025-04-02T02%3A39%3A01Z&se=2025-04-08T03%3A39%3A01Z&sks=b&skt=2025-04-02T02%3A39%3A01Z&ske=2025-04-08T03%3A39%3A01Z&sktid=a48cca56-e6da-484e-a814-9c849652bcb3&skoid=3d249c53-07fa-4ba4-9b65-0bf8eb4ea46a&skv=2019-02-02&sv=2018-11-09&sr=b&sp=r&spr=https%2Chttp&sig=9aU9xKI0BkhRtrTb0RD05%2BT6prBnmXk1I7nV2nTsFbY%3D&az=oaivgprodscus"
+}
+*/
+
+#[derive(Debug, Deserialize)]
+pub struct SoraMediaUploadResponse {
+  /// This is the media token id.
+  pub id: String,
+  pub r#type: String,
+  pub created_at: String,
+  pub filename: String,
+  pub extension: String,
+  pub mime_type: String,
+  pub url: String,
+  pub width: Option<u32>,
+  pub height: Option<u32>,
+  pub duration_sec: Option<f64>,
+  pub n_frames: Option<u32>,
+  pub size_bytes: u64,
+  pub thumbnail_url: Option<String>,
+}
+
+pub (crate) struct SoraMediaUploadRequest<'a> {
+  pub file_path: String,
+  pub credentials: &'a SoraCredentials,
+}
+
+pub (crate) async fn upload_media_http_request(file_bytes: Vec<u8>, filename: String, mime_type: &str, credentials: &SoraCredentials) -> AnyhowResult<SoraMediaUploadResponse> {
+  // Create multipart form
+  let part = Part::bytes(file_bytes) // NB: Reqwest needs to own the bytes.
+    .file_name(filename) // NB: Reqwest needs to own the bytes
+    .mime_str(mime_type)?;
+
+  let form = Form::new().part("file", part);
+
+  // Make API request
+  let client = Client::new();
+  let request_builder = client.post(SORA_UPLOAD_MEDIA_URL)
+      .multipart(form)
+      .header("User-Agent", USER_AGENT)
+      .header("Cookie", &credentials.cookie)
+      .header("Authorization", credentials.authorization_header_value());
+
+  let response = request_builder.send().await?;
+
+  // Check response status
+  if !response.status().is_success() {
+    let status = response.status();
+    let text = response.text().await?;
+    return Err(anyhow::anyhow!("Upload failed with status {}: {}", status, text));
+  }
+
+  // Parse response
+  let upload_response = response.json::<SoraMediaUploadResponse>().await?;
+  Ok(upload_response)
+}

crates/lib/image_algorithms/Cargo.toml

@@ -0,0 +1,30 @@
+[package]
+name = "image_algorithms"
+edition = "2018"
+version = "0.0.1"
+authors = [
+    "Brandon Thomas <bt@brand.io>",
+    "Brandon Thomas <echelon@gmail.com>",
+]
+publish = false
+
+[lib]
+name = "image_algorithms"
+path = "src/lib.rs"
+
+[package.metadata.cargo-udeps.ignore]
+normal = ["my-workspace-hack"]
+
+[dependencies]
+
+# Workspace hack (faster compile times)
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }
+
+# Internal
+errors = { path = "../errors" }
+
+# External
+image = "0.25.6"
+
+[dev-dependencies]
+testing = { path = "../testing" }

crates/lib/image_algorithms/src/lib.rs

@@ -0,0 +1 @@
+pub mod remove_image_border;

crates/lib/image_algorithms/src/remove_image_border.rs

@@ -0,0 +1,78 @@
+use image::{DynamicImage, Pixel};
+use errors::AnyhowResult;
+
+pub fn remove_image_border(image: &DynamicImage) -> AnyhowResult<DynamicImage> {
+  let mut image = image.to_rgba8();
+
+  let mut top = 0u32;
+  let mut bottom = image.height();
+  let mut left = 0u32;
+  let mut right = image.width();
+
+  // TODO: Iterate with binary search - it'll probably be much faster.
+
+  for (i, mut row) in image.rows().enumerate() {
+    let count_non_transparent = row.filter(|pixel| pixel.0[3] != 0).count();
+    if count_non_transparent < 5 {
+      top = i as u32;
+    } else {
+      break;
+    }
+  }
+
+  for (i, mut row) in image.rows().enumerate().rev() {
+    //let all_transparent = row.all(|pixel| pixel.0[3] == 0);
+    let count_non_transparent = row.filter(|pixel| pixel.0[3] != 0).count();
+    //if all_transparent {
+    if count_non_transparent < 5 {
+      bottom = i as u32;
+    } else {
+      break;
+    }
+  }
+
+  for (x, y, pixel) in image.enumerate_pixels() {
+    
+  }
+
+  for (y, mut row) in image.enumerate_rows() {
+    for (x, y, pixel) in row {
+      
+    }
+  }
+
+
+  let image = DynamicImage::ImageRgba8(image);
+
+  let width = image.width();
+  let height = bottom - top;
+
+  println!("top: {},  bottom: {}", top, bottom);
+  println!("orig width: {},  height: {}", image.width(), image.height());
+  println!("new width: {},  height: {}", width, height);
+
+  let image = image.crop_imm(0, top, width, height);
+
+  Ok(image)
+}
+
+#[cfg(test)]
+mod tests {
+  use image::{ImageFormat, ImageReader};
+  use errors::AnyhowResult;
+  use testing::test_file_path::test_file_path;
+  use crate::remove_image_border::remove_image_border;
+
+  #[ignore] // NB: Not going to check these files into the repo. Just for manual testing.
+  #[test]
+  fn test_crop_image() -> AnyhowResult<()> {
+    let filename = test_file_path("test_data/image/mario.png")?;
+    let image = ImageReader::open(filename)?.decode()?;
+
+    let image = remove_image_border(&image)?;
+
+    image.save_with_format("testing_output.png", ImageFormat::Png)?;
+
+    Ok(())
+  }
+}

crates/lib/mysql_queries/src/payloads/generic_inference_args/generic_inference_args.rs

@@ -1,7 +1,6 @@
 use crate::payloads::generic_inference_args::inner_payloads::f5_tts_payload::F5TTSPayload;
 use crate::payloads::generic_inference_args::inner_payloads::face_fusion_payload::FaceFusionPayload;
 use crate::payloads::generic_inference_args::inner_payloads::gptsovits_payload::GptSovitsPayload;
-use crate::payloads::generic_inference_args::inner_payloads::image_generation_payload::StableDiffusionArgs;
 use crate::payloads::generic_inference_args::inner_payloads::lipsync_payload::LipsyncArgs;
 use crate::payloads::generic_inference_args::inner_payloads::live_portrait_payload::LivePortraitPayload;
 use crate::payloads::generic_inference_args::inner_payloads::mocap_payload::MocapArgs;
@@ -12,6 +11,8 @@ use crate::payloads::generic_inference_args::inner_payloads::tts_payload::TTSArg
 use crate::payloads::generic_inference_args::inner_payloads::videofilter_payload::RerenderArgs;
 use crate::payloads::generic_inference_args::inner_payloads::workflow_payload::WorkflowArgs;
 use errors::AnyhowResult;
+use crate::payloads::generic_inference_args::inner_payloads::image_generation_payload::StableDiffusionArgs;
+use crate::payloads::generic_inference_args::inner_payloads::sora_image_gen_args::SoraImageGenArgs;
 
 /// Used to encode extra state for the `generic_inference_jobs` table.
 /// This should act somewhat like a serialized protobuf stored inside a record.
@@ -111,7 +112,7 @@ pub enum PolymorphicInferenceArgs {
   Tts(TTSArgs),
     // No arguments yet.
     // It might be best to just not include this when not used.
-    
+
   /// Voice conversion. (Short name to save space when serializing.)
   Vc {
     /// Argument for so-vits-svc
@@ -143,6 +144,10 @@ pub enum PolymorphicInferenceArgs {
   /// Image generation. (Short name to save space when serializing.)
   Ig(StableDiffusionArgs),
 
+
+  /// Sora Image Generation. (Short name to save space when serializing.)
+  Sg(SoraImageGenArgs),
+
   /// Mocap (Short name to save space when serializing.)
   Mc(MocapArgs),
 

crates/lib/mysql_queries/src/payloads/generic_inference_args/inner_payloads/image_generation_payload.rs

@@ -1,3 +1,4 @@
+use tokens::tokens::media_files::MediaFileToken;
 use tokens::tokens::model_weights::ModelWeightToken;
 
 /// Video sources can be one of several:
@@ -52,7 +53,7 @@ pub struct StableDiffusionArgs {
     #[serde(rename = "se")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub maybe_seed: Option<i64>,
-  
+
     #[serde(rename = "mu")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub maybe_upload_path: Option<String>,
@@ -60,7 +61,7 @@ pub struct StableDiffusionArgs {
     #[serde(rename = "cf")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub maybe_cfg_scale: Option<u32>,
-    
+
     #[serde(rename = "lu")]
     #[serde(skip_serializing_if = "Option::is_none")]
     pub maybe_lora_upload_path: Option<String>,

crates/lib/mysql_queries/src/payloads/generic_inference_args/inner_payloads/mod.rs

@@ -1,3 +1,4 @@
+pub mod f5_tts_payload;
 pub mod face_fusion_payload;
 pub mod gptsovits_payload;
 pub mod image_generation_payload;
@@ -5,9 +6,9 @@ pub mod lipsync_payload;
 pub mod live_portrait_payload;
 pub mod mocap_payload;
 pub mod render_engine_scene_to_video_payload;
+pub mod seed_vc_payload;
+pub mod sora_image_gen_args;
 pub mod studio_gen2_payload;
 pub mod tts_payload;
 pub mod videofilter_payload;
 pub mod workflow_payload;
-pub mod f5_tts_payload;
-pub mod seed_vc_payload;

crates/lib/mysql_queries/src/payloads/generic_inference_args/inner_payloads/sora_image_gen_args.rs

@@ -0,0 +1,23 @@
+use tokens::tokens::media_files::MediaFileToken;
+
+#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
+pub struct SoraImageGenArgs {
+  #[serde(rename = "p")]
+  #[serde(skip_serializing_if = "Option::is_none")]
+  pub prompt: Option<String>,
+
+  #[serde(rename = "sm")]
+  pub scene_snapshot_media_token: Option<MediaFileToken>,
+
+  #[serde(rename = "mf")]
+  #[serde(skip_serializing_if = "Option::is_none")]
+  pub maybe_additional_media_file_tokens: Option<Vec<MediaFileToken>>,
+
+  #[serde(rename = "sa")]
+  #[serde(skip_serializing_if = "Option::is_none")]
+  pub maybe_number_of_samples: Option<u32>,
+
+  #[serde(rename = "su")]
+  #[serde(skip_serializing_if = "Option::is_none")]
+  pub maybe_sora_media_upload_tokens: Option<Vec<String>>,
+}

crates/lib/testing/src/test_file_path.rs

@@ -17,7 +17,9 @@ pub fn test_file_path<P: AsRef<Path>>(path_from_repo_root: P) -> AnyhowResult<Pa
       .to_str()
       .ok_or(anyhow!("path cannot convert to string"))?;
   path.push(format!("../../../{}", file));
+  println!(" >>> path = {:?}", path);
   Ok(path.canonicalize()?)
+
 }
 
 #[cfg(test)]

crates/service/web/storyteller_web/Cargo.toml

@@ -69,6 +69,7 @@ tts_common = { path = "../../../lib/tts_common" }
 url_config = { path = "../../../lib/url_config" }
 user_input_common = { path = "../../../lib/user_input_common" }
 videos = { path = "../../../lib/files/videos" }
+openai_sora_client = { path = "../../../lib/api_clients/openai_sora_client" }
 
 # Internal Components
 billing_component = { path = "../../plugins/billing" }

crates/service/web/storyteller_web/src/docs/api_doc.rs

@@ -124,6 +124,8 @@ use crate::http_server::endpoints::stats::get_unified_queue_stats_handler::*;
 use crate::http_server::endpoints::tags::list_tags_for_entity_handler::*;
 use crate::http_server::endpoints::tags::set_tags_for_entity_handler::*;
 use crate::http_server::endpoints::tts::enqueue_infer_tts_handler::enqueue_infer_tts_handler::*;
+use crate::http_server::endpoints::image_studio::prompt::enqueue_studio_image_generation_handler::*;
+use crate::http_server::endpoints::image_studio::upload::upload_snapshot_media_file_handler::*;
 use crate::http_server::endpoints::user_bookmarks::batch_get_user_bookmarks_handler::*;
 use crate::http_server::endpoints::user_bookmarks::create_user_bookmark_handler::*;
 use crate::http_server::endpoints::user_bookmarks::delete_user_bookmark_handler::*;
@@ -207,6 +209,8 @@ use crate::http_server::web_utils::response_success_helpers::*;
     crate::http_server::endpoints::media_files::upload::upload_engine_asset::upload_engine_asset_media_file_handler::upload_engine_asset_media_file_handler,
     crate::http_server::endpoints::media_files::upload::upload_generic::upload_media_file_handler::upload_media_file_handler,
     crate::http_server::endpoints::media_files::upload::upload_image_media_file_handler::upload_image_media_file_handler,
+    crate::http_server::endpoints::image_studio::prompt::enqueue_studio_image_generation_handler::enqueue_studio_image_generation_handler,
+    crate::http_server::endpoints::image_studio::upload::upload_snapshot_media_file_handler::upload_snapshot_media_file_handler,
     crate::http_server::endpoints::studio_gen2::enqueue_studio_gen2_handler::enqueue_studio_gen2_handler,
     crate::http_server::endpoints::media_files::upload::upload_new_engine_asset_media_file_handler::upload_new_engine_asset_media_file_handler,
     crate::http_server::endpoints::media_files::upload::upload_new_scene_media_file_handler::upload_new_scene_media_file_handler,
@@ -336,6 +340,9 @@ use crate::http_server::web_utils::response_success_helpers::*;
     AppStateServerInfo,
     AppStateStatusAlertCategory,
     AppStateStatusAlertInfo,
+    EnqueueStudioImageGenRequest,
+    EnqueueImageGenRequestSuccessResponse,
+    EnqueueImageGenRequestError,
     AppStateSubscriptionProductKey,
     AppStateUserInfo,
     AppStateUserLocale,
@@ -347,6 +354,8 @@ use crate::http_server::web_utils::response_success_helpers::*;
     BatchGetMediaFilesQueryParams,
     BatchGetMediaFilesSuccessResponse,
     BatchGetUserBookmarksError,
+    UploadSnapshotMediaFileForm,
+    UploadSnapshotMediaFileSuccessResponse,
     BatchGetUserBookmarksQueryParams,
     BatchGetUserBookmarksResponse,
     BatchGetUserRatingError,

crates/service/web/storyteller_web/src/http_server/endpoints/control_plane/mod.rs

@@ -0,0 +1 @@
+pub mod set_sora_secret_handler;

crates/service/web/storyteller_web/src/http_server/endpoints/control_plane/set_sora_secret_handler.rs

@@ -0,0 +1,137 @@
+use std::fmt;
+use std::sync::Arc;
+
+use actix_web::{HttpRequest, HttpResponse, web, Either};
+use actix_web::error::ResponseError;
+use actix_web::http::StatusCode;
+use actix_web::web::{Form, Json, Path};
+use anyhow::anyhow;
+use log::{error, log, warn};
+use r2d2_redis::redis::Commands;
+use utoipa::ToSchema;
+
+use enums::by_table::media_files::media_file_animation_type::MediaFileAnimationType;
+use enums::by_table::media_files::media_file_engine_category::MediaFileEngineCategory;
+use http_server_common::response::serialize_as_json_error::serialize_as_json_error;
+use mysql_queries::queries::media_files::edit::update_media_file_animation_type::update_media_file_animation_type;
+use mysql_queries::queries::media_files::get::get_media_file::get_media_file;
+use tokens::tokens::media_files::MediaFileToken;
+
+use crate::state::server_state::ServerState;
+use crate::util::redis_sora_secrets::get_sora_credentials;
+use crate::util::redis_sora_secrets::set_sora_credential_field;
+use crate::util::redis_sora_secrets::RedisSoraCredentialSubkey;
+
+#[derive(Deserialize, Clone, Copy)]
+#[serde(rename_all = "snake_case")]
+pub enum SetSoraSecretType {
+  Bearer,
+  Cookie,
+  Sentinel
+}
+
+#[derive(Deserialize, ToSchema)]
+pub struct SetSoraSecretRequest {
+  pub key: SetSoraSecretType,
+  pub value: String,
+}
+
+#[derive(Serialize, ToSchema)]
+pub struct SetSoraSecretSuccessResponse {
+  pub success: bool,
+  pub bearer: String,
+  pub cookie: String,
+  pub sentinel: String,
+}
+
+// =============== Error Response ===============
+
+#[derive(Debug, Serialize, ToSchema)]
+pub enum SetSoraSecretError {
+  BadInput(String),
+  NotAuthorized,
+  ServerError,
+}
+
+impl ResponseError for SetSoraSecretError {
+  fn status_code(&self) -> StatusCode {
+    match *self {
+      SetSoraSecretError::BadInput(_) => StatusCode::BAD_REQUEST,
+      SetSoraSecretError::NotAuthorized => StatusCode::UNAUTHORIZED,
+      SetSoraSecretError::ServerError => StatusCode::INTERNAL_SERVER_ERROR,
+    }
+  }
+
+  fn error_response(&self) -> HttpResponse {
+    serialize_as_json_error(self)
+  }
+}
+
+// NB: Not using derive_more::Display since Clion doesn't understand it.
+impl fmt::Display for SetSoraSecretError {
+  fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+    write!(f, "{:?}", self)
+  }
+}
+
+// =============== Handler ===============
+
+/// Set a Sora secret
+#[utoipa::path(
+  post,
+  tag = "Control Plane",
+  path = "/v1/control_plane/sora_secret",
+  responses(
+        (status = 200, description = "Success", body = SetSoraSecretSuccessResponse),
+        (status = 400, description = "Bad input", body = SetSoraSecretError),
+        (status = 401, description = "Not authorized", body = SetSoraSecretError),
+        (status = 500, description = "Server error", body = SetSoraSecretError),
+  ),
+  params(
+        ("request" = SetSoraSecretRequest, description = "Payload for Request"),
+        ("path" = MediaFileTokenPathInfo, description = "Path for Request")
+  )
+)]
+pub async fn set_sora_secret_handler(
+  http_request: HttpRequest,
+  request: Form<SetSoraSecretRequest>,
+  server_state: web::Data<Arc<ServerState>>
+) -> Result<Json<SetSoraSecretSuccessResponse>, SetSoraSecretError> {
+
+  // TODO(bt,2025-04-03): Some form of authentication should be required for this endpoint.
+  let value = request.value.trim().to_string();
+
+  let mut redis = server_state.redis_pool
+      .get()
+      .map_err(|e| {
+        error!("redis error: {:?}", e);
+        SetSoraSecretError::ServerError
+      })?;
+
+  set_sora_credential_field(
+    &mut redis,
+    match request.key {
+      SetSoraSecretType::Bearer => RedisSoraCredentialSubkey::Bearer,
+      SetSoraSecretType::Cookie => RedisSoraCredentialSubkey::Cookie,
+      SetSoraSecretType::Sentinel => RedisSoraCredentialSubkey::Sentinel,
+    },
+    &value
+  ).map_err(|e| {
+    error!("redis error: {:?}", e);
+    SetSoraSecretError::ServerError
+  })?;
+
+  let credentials = get_sora_credentials(&mut redis)
+      .map_err(|e| {
+        error!("redis error: {:?}", e);
+        SetSoraSecretError::ServerError
+      })?;
+
+  // TODO(bt,2025-04-03): Probably not smart to return, but useful for debugging.
+  Ok(Json(SetSoraSecretSuccessResponse {
+    success: true,
+    bearer: credentials.bearer_token,
+    cookie: credentials.cookie,
+    sentinel: credentials.sentinel.unwrap_or_else(|| "".to_string()),
+  }))
+}

crates/service/web/storyteller_web/src/http_server/endpoints/image_gen/enqueue_image_generation.rs

@@ -507,6 +507,9 @@ fn validate_request(
     Ok(())
 }
 
+
+
+
 fn inference_mode_from_http_request(http_request: &HttpRequest) -> Option<TypeOfInference> {
     inference_mode_from_url_path(http_request.path())
 }

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/get_sora_credentials_from_request.rs

@@ -0,0 +1,68 @@
+use actix_web::HttpRequest;
+use anyhow::anyhow;
+use log::{error, info};
+use r2d2_redis::r2d2::PooledConnection;
+use r2d2_redis::RedisConnectionManager;
+use errors::AnyhowResult;
+use http_server_common::request::get_request_header_optional::get_request_header_optional;
+use openai_sora_client::credentials::SoraCredentials;
+use crate::util::redis_sora_secrets::get_sora_credentials;
+
+const SORA_BEARER_HEADER_NAME : &str = "sora-bearer";
+const SORA_COOKIE_HEADER_NAME : &str = "sora-cookie";
+const SORA_SENTINEL_HEADER_NAME : &str = "sora-sentinel";
+
+pub fn get_sora_credentials_from_request(
+  http_request: &HttpRequest,
+  redis: &mut PooledConnection<RedisConnectionManager>,
+) -> AnyhowResult<SoraCredentials> {
+
+  let sora_bearer = get_request_header_optional(&http_request, SORA_BEARER_HEADER_NAME);
+  let sora_cookie = get_request_header_optional(&http_request, SORA_COOKIE_HEADER_NAME);
+  let sora_sentinel = get_request_header_optional(&http_request, SORA_SENTINEL_HEADER_NAME);
+
+  match (&sora_bearer, &sora_cookie, &sora_sentinel) {
+    (Some(bearer), Some(cookie), Some(sentinel)) => {
+      info!("Sora credentials were present in HTTP request headers!");
+
+      return Ok(SoraCredentials {
+        bearer_token: bearer.to_string(),
+        cookie: cookie.to_string(),
+        sentinel: Some(sentinel.to_string()),
+      });
+    }
+    _ => {} // Fall through
+  }
+
+  let redis_result = get_sora_credentials(redis);
+
+  match redis_result {
+    Ok(mut credentials) => {
+      if let Some(bearer) = sora_bearer {
+        info!("override bearer value");
+        credentials.bearer_token = bearer;
+      }
+      if let Some(cookie) = sora_cookie {
+        info!("override cookie value");
+        credentials.cookie = cookie;
+      }
+      if let Some(sentinel) = sora_sentinel {
+        info!("override sentinel value");
+        credentials.sentinel = Some(sentinel);
+      }
+      return Ok(credentials)
+    },
+    Err(err) => {
+      error!("Failed to get Sora credentials from Redis; using env as backup. Error: {:?}", err);
+      // NB: Fall through
+    }
+  };
+
+  match SoraCredentials::from_env() {
+    Ok(creds) => Ok(creds),
+    Err(err) => {
+      error!("Failed to load Sora credentials from environment: {:?}", err);
+      Err(anyhow!("Failed to load Sora credentials from environment"))
+    }
+  }
+}

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/mod.rs

@@ -0,0 +1,3 @@
+pub mod upload;
+pub mod prompt;
+pub mod get_sora_credentials_from_request;

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/prompt/enqueue_studio_image_generation_handler.rs

@@ -0,0 +1,465 @@
+#![forbid(unused_mut)]
+
+use std::fmt::Debug;
+use std::fmt::{Display, Formatter};
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use actix_web::error::ResponseError;
+use actix_web::http::StatusCode;
+use actix_web::web::Json;
+use actix_web::{web, HttpRequest, HttpResponse};
+use bucket_paths::legacy::typified_paths::public::media_files::bucket_file_path::MediaFileBucketPath;
+use cloud_storage::bucket_client::BucketClient;
+use log::{debug, error, info};
+use log::warn;
+use mysql_queries::queries::media_files::get::get_media_file::get_media_file;
+use openai_sora_client::credentials::SoraCredentials;
+use serde::Deserialize;
+use serde::Serialize;
+use sqlx::MySqlPool;
+use tempdir::TempDir;
+use utoipa::ToSchema;
+use web::Data;
+
+use enums::by_table::generic_inference_jobs::inference_category::InferenceCategory;
+use enums::by_table::generic_inference_jobs::inference_job_type::InferenceJobType;
+use enums::by_table::generic_inference_jobs::inference_model_type::InferenceModelType;
+use enums::common::visibility::Visibility;
+use http_server_common::request::get_request_header_optional::get_request_header_optional;
+use http_server_common::request::get_request_ip::get_request_ip;
+use mysql_queries::payloads::generic_inference_args::generic_inference_args::{
+  GenericInferenceArgs,
+  InferenceCategoryAbbreviated,
+  PolymorphicInferenceArgs,
+};
+use mysql_queries::payloads::generic_inference_args::inner_payloads::sora_image_gen_args::SoraImageGenArgs;
+use mysql_queries::queries::generic_inference::web::insert_generic_inference_job::{
+  insert_generic_inference_job,
+  InsertGenericInferenceArgs,
+};
+use mysql_queries::queries::idepotency_tokens::insert_idempotency_token::insert_idempotency_token;
+use openai_sora_client::image_gen::common::{ImageSize, NumImages};
+use openai_sora_client::image_gen::sora_image_gen_remix::{sora_image_gen_remix, SoraImageGenRemixRequest};
+use openai_sora_client::upload::upload_media_from_file::{sora_media_upload_from_file, SoraMediaUploadRequest};
+use tokens::tokens::generic_inference_jobs::InferenceJobToken;
+use tokens::tokens::media_files::MediaFileToken;
+use tokens::tokens::model_weights::ModelWeightToken;
+use tokens::tokens::users::UserToken;
+
+use crate::configs::plans::get_correct_plan_for_session::get_correct_plan_for_session;
+use crate::http_server::endpoints::image_studio::get_sora_credentials_from_request::get_sora_credentials_from_request;
+use crate::http_server::endpoints::tts::get_tts_inference_job_status::GetTtsInferenceStatusError;
+use crate::http_server::validations::validate_idempotency_token_format::validate_idempotency_token_format;
+use crate::http_server::web_utils::response_error_helpers::to_simple_json_error;
+use crate::state::server_state::ServerState;
+use crate::util::allowed_studio_access::allowed_studio_access;
+use crate::util::redis_sora_secrets::get_sora_credentials;
+
+/// This is the number of images (batch size) to generate for each request.
+/// We should allow all users to have multiple images generated at once as this
+/// is what other providers do.
+const MINIMUM_IMAGE_COUNT : u32 = 1;
+
+/// The maximum number of images (batch size) to generate for each request.
+const MAXIMUM_IMAGE_COUNT : u32 = 2;
+
+/// Debug requests can get routed to special "debug-only" workers, which can
+/// be used to trial new code, run debugging, etc.
+const DEBUG_HEADER_NAME: &str = "enable-debug-mode";
+
+/// The routing tag header can send workloads to particular k8s hosts.
+/// This is useful for catching the live logs or intercepting the job.
+const ROUTING_TAG_HEADER_NAME: &str = "routing-tag";
+
+#[derive(Deserialize, ToSchema)]
+pub struct EnqueueStudioImageGenRequest {
+  pub uuid_idempotency_token: String,
+
+  /// Image media file; the engine or canvas snapshot (screenshot).
+  pub snapshot_media_token: MediaFileToken,
+
+  /// The user's image generation prompt.
+  pub prompt: String,
+
+  /// Additional images to include (optional). Up to nine images.
+  pub maybe_additional_images: Option<Vec<MediaFileToken>>,
+
+  pub maybe_number_of_samples: Option<u32>,
+}
+
+#[derive(Serialize, ToSchema)]
+pub struct EnqueueImageGenRequestSuccessResponse {
+  pub success: bool,
+  pub inference_job_token: InferenceJobToken,
+}
+
+#[derive(Debug, ToSchema)]
+pub enum EnqueueImageGenRequestError {
+  BadInput(String),
+  NotAuthorized,
+  ServerError,
+  RateLimited,
+}
+
+impl ResponseError for EnqueueImageGenRequestError {
+  fn status_code(&self) -> StatusCode {
+    match *self {
+      EnqueueImageGenRequestError::BadInput(_) => StatusCode::BAD_REQUEST,
+      EnqueueImageGenRequestError::NotAuthorized => StatusCode::UNAUTHORIZED,
+      EnqueueImageGenRequestError::ServerError => StatusCode::INTERNAL_SERVER_ERROR,
+      EnqueueImageGenRequestError::RateLimited => StatusCode::TOO_MANY_REQUESTS,
+    }
+  }
+
+  fn error_response(&self) -> HttpResponse {
+    let error_reason = match self {
+      EnqueueImageGenRequestError::BadInput(reason) => reason.to_string(),
+      EnqueueImageGenRequestError::NotAuthorized => "unauthorized".to_string(),
+      EnqueueImageGenRequestError::ServerError => "server error".to_string(),
+      EnqueueImageGenRequestError::RateLimited => "rate limited".to_string(),
+    };
+
+    to_simple_json_error(&error_reason, self.status_code())
+  }
+}
+
+impl Display for EnqueueImageGenRequestError {
+  fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+    write!(f, "{:?}", self)
+  }
+}
+
+/// Prompt image generation using image studio
+#[utoipa::path(
+  post,
+  tag = "Image Studio",
+  path = "/v1/image_studio/prompt",
+  responses(
+        (
+            status = 200,
+            description = "Enqueue TTS generically",
+            body = EnqueueImageGenRequestSuccessResponse,
+        ),
+        (status = 400, description = "Bad input", body = EnqueueImageGenRequestError),
+        (status = 401, description = "Not authorized", body = EnqueueImageGenRequestError),
+        (status = 429, description = "Rate limited", body = EnqueueImageGenRequestError),
+        (status = 500, description = "Server error", body = EnqueueImageGenRequestError)
+  ),
+  params(("request" = EnqueueStudioImageGenRequest, description = "Payload for Image Generation Request"))
+)]
+pub async fn enqueue_studio_image_generation_handler(
+  http_request: HttpRequest,
+  request: Json<EnqueueStudioImageGenRequest>,
+  server_state: Data<Arc<ServerState>>
+) -> Result<HttpResponse, EnqueueImageGenRequestError> {
+
+  validate_request(&request)?;
+
+  let mut mysql_connection = server_state.mysql_pool.acquire().await.map_err(|err| {
+    warn!("MySql pool error: {:?}", err);
+    EnqueueImageGenRequestError::ServerError
+  })?;
+
+  // ==================== USER SESSION ==================== //
+
+  let maybe_user_session = server_state.session_checker
+    .maybe_get_user_session_extended_from_connection(&http_request, &mut mysql_connection).await
+    .map_err(|e| {
+      warn!("Session checker error: {:?}", e);
+      EnqueueImageGenRequestError::ServerError
+    })?;
+
+  let mut maybe_user_token: Option<UserToken> = None;
+
+  if let Some(user_session) = maybe_user_session.as_ref() {
+    maybe_user_token = Some(UserToken::new_from_str(&user_session.user_token));
+  }
+
+  // ==================== PLANS ==================== //
+
+  // Plan should handle "first anonymous use" and "investor" cases.
+  let plan = get_correct_plan_for_session(
+    server_state.server_environment_old,
+    maybe_user_session.as_ref()
+  );
+
+  // Separate priority for animation.
+  let priority_level = plan.web_vc_base_priority_level();
+
+  // ==================== DEBUG MODE + ROUTING TAG ==================== //
+
+  let is_debug_request = get_request_header_optional(&http_request, DEBUG_HEADER_NAME).is_some();
+
+  let maybe_routing_tag = get_request_header_optional(&http_request, ROUTING_TAG_HEADER_NAME).map(
+    |routing_tag| routing_tag.trim().to_string()
+  );
+
+  // ==================== BANNED USERS ==================== //
+
+  if let Some(ref user) = maybe_user_session {
+    if user.role.is_banned {
+      return Err(EnqueueImageGenRequestError::NotAuthorized);
+    }
+  }
+
+  // DETECT premium user and queue
+
+  // ==================== RATE LIMIT ==================== //
+
+  let rate_limiter = match maybe_user_session {
+    None => &server_state.redis_rate_limiters.logged_out,
+    Some(ref _user) => &server_state.redis_rate_limiters.logged_in,
+  };
+
+  if let Err(_err) = rate_limiter.rate_limit_request(&http_request) {
+    return Err(EnqueueImageGenRequestError::RateLimited);
+  }
+
+  // Get up IP address
+  let ip_address = get_request_ip(&http_request);
+
+  // ==================== HANDLE IDEMPOTENCY ==================== //
+
+  if let Err(reason) = validate_idempotency_token_format(&request.uuid_idempotency_token) {
+    return Err(EnqueueImageGenRequestError::BadInput(reason));
+  }
+
+  insert_idempotency_token(&request.uuid_idempotency_token, &mut *mysql_connection)
+    .await
+    .map_err(|err| {
+      error!("Error inserting idempotency token: {:?}", err);
+      EnqueueImageGenRequestError::BadInput("invalid idempotency token".to_string())
+    })?;
+
+  // ==================== DOWNLOAD FILES AND UPLOAD TO SORA ==================== //
+
+  // TODO: Maybe this moves to a job.
+
+  let work_temp_dir = server_state.temp_dir_creator.new_tempdir("image_studio").unwrap();
+
+  let public_bucket_client = server_state.public_bucket_client.clone();
+
+  let scene_media_token = request.snapshot_media_token.clone();
+
+  let scene_media_path = query_and_download_media_file(
+    &scene_media_token,
+    &work_temp_dir,
+    &public_bucket_client,
+    &server_state.mysql_pool
+  ).await.map_err(|err| {
+    error!("Failed to download scene media file: {:?}", err);
+    EnqueueImageGenRequestError::ServerError
+  })?;
+
+  let mut files_to_upload = Vec::with_capacity(1 + request.maybe_additional_images
+      .as_ref().map(|v| v.len()).unwrap_or(0));
+
+  files_to_upload.push(scene_media_path.clone());
+
+  let additional_images = request.maybe_additional_images.clone().unwrap_or_else(|| vec![]);
+
+  for media_file_token in &additional_images {
+    let media_file_path = query_and_download_media_file(
+      &media_file_token,
+      &work_temp_dir,
+      &public_bucket_client,
+      &server_state.mysql_pool
+    ).await.map_err(|err| {
+      error!("Failed to download additional media file: {:?}", err);
+      EnqueueImageGenRequestError::ServerError
+    })?;
+
+    files_to_upload.push(media_file_path.clone());
+  }
+
+  // ==================== HANDLE SORA CREDENTIALS ==================== //
+
+  let mut redis = server_state.redis_pool
+      .get()
+      .map_err(|e| {
+        error!("redis error: {:?}", e);
+        EnqueueImageGenRequestError::ServerError
+      })?;
+
+  let sora_credentials = get_sora_credentials_from_request(&http_request, &mut redis)
+      .map_err(|e| {
+        error!("sora credential error: {:?}", e);
+        EnqueueImageGenRequestError::ServerError
+      })?;
+
+  // ==================== HANDLE SORA UPLOAD ==================== //
+
+  let mut sora_media_tokens = Vec::with_capacity(files_to_upload.len());
+
+  for file_path in files_to_upload {
+    let scene_media_upload_request = SoraMediaUploadRequest {
+      file_path: scene_media_path.to_string_lossy().to_string(),
+      credentials: &sora_credentials,
+    };
+
+    let sora_upload_response = sora_media_upload_from_file(file_path, &sora_credentials)
+        .await
+        .map_err(|err| {
+          error!("Failed to upload scene media to Sora: {:?}", err);
+          EnqueueImageGenRequestError::ServerError
+        })?;
+
+    debug!("Uploaded media to Sora : {:?}", sora_upload_response);
+    sora_media_tokens.push(sora_upload_response.id);
+  }
+
+  // ==================== HANDLE SORA PROMPT ==================== //
+
+  let number_of_samples = match request.maybe_number_of_samples {
+    Some(val) => if val > MAXIMUM_IMAGE_COUNT {
+      MAXIMUM_IMAGE_COUNT
+    } else if val < MINIMUM_IMAGE_COUNT {
+      MINIMUM_IMAGE_COUNT
+    } else {
+      val
+    },
+    None => 1,
+  };
+
+  let prompt = request.prompt.clone();
+
+  let response = sora_image_gen_remix(SoraImageGenRemixRequest {
+    prompt: prompt.clone(),
+    num_images: NumImages::One,
+    image_size: ImageSize::Square,
+    sora_media_tokens: sora_media_tokens.clone(),
+    credentials: &sora_credentials,
+  }).await
+    .map_err(|err| {
+      error!("Failed to call Sora image generation: {:?}", err);
+      EnqueueImageGenRequestError::ServerError
+    })?;
+
+  debug!("Sora image gen response: {:?}", response);
+
+  // Store the actual inference args that will go to the database
+  let inference_args = SoraImageGenArgs {
+    prompt: Some(prompt),
+    scene_snapshot_media_token: Some(request.snapshot_media_token.clone()),
+    maybe_additional_media_file_tokens: Some(additional_images),
+    maybe_number_of_samples: Some(number_of_samples),
+    maybe_sora_media_upload_tokens: Some(sora_media_tokens),
+  };
+
+  // create the inference args here
+  let maybe_avt_token = server_state.avt_cookie_manager.get_avt_token_from_request(&http_request);
+
+  // create the job record here!
+  let query_result = insert_generic_inference_job(InsertGenericInferenceArgs {
+    uuid_idempotency_token: &request.uuid_idempotency_token,
+    job_type: InferenceJobType::ImageGenApi,
+    maybe_product_category: None, // This is not a product anymore
+    inference_category: InferenceCategory::ImageGeneration,
+    maybe_model_type: Some(InferenceModelType::ImageGenApi), // NB: Model is static during inference
+    maybe_model_token: None, // NB: Model is static during inference
+    maybe_input_source_token: None,
+    maybe_input_source_token_type: None,
+    maybe_download_url: None,
+    maybe_cover_image_media_file_token: None,
+    maybe_raw_inference_text: None,
+    maybe_max_duration_seconds: None,
+    maybe_inference_args: Some(GenericInferenceArgs {
+      inference_category: Some(InferenceCategoryAbbreviated::ImageGeneration),
+      args: Some(PolymorphicInferenceArgs::Sg(inference_args)),
+    }),
+    maybe_creator_user_token: maybe_user_token.as_ref(),
+    maybe_avt_token: maybe_avt_token.as_ref(),
+    creator_ip_address: &ip_address,
+    creator_set_visibility: Visibility::Public,
+    priority_level,
+    requires_keepalive: false, //reverse ...  TODO fix this. we set it base on account is premium or not ...
+    is_debug_request,
+    maybe_routing_tag: maybe_routing_tag.as_deref(),
+    mysql_pool: &server_state.mysql_pool,
+  }).await;
+
+  let job_token = match query_result {
+    Ok((job_token, _id)) => job_token,
+    Err(err) => {
+      warn!("New generic inference job creation DB error: {:?}", err);
+      if err.had_duplicate_idempotency_token() {
+        return Err(EnqueueImageGenRequestError::BadInput("Duplicate idempotency token".to_string()));
+      }
+      return Err(EnqueueImageGenRequestError::ServerError);
+    }
+  };
+
+  let response: EnqueueImageGenRequestSuccessResponse = EnqueueImageGenRequestSuccessResponse {
+    success: true,
+    inference_job_token: job_token,
+  };
+
+  let body = serde_json::to_string(&response)
+    .map_err(|_e| EnqueueImageGenRequestError::ServerError)?;
+
+  // Error handling 101 rust result type returned like so.
+  Ok(HttpResponse::Ok().content_type("application/json").body(body))
+}
+
+fn validate_request(
+  request: &Json<EnqueueStudioImageGenRequest>,
+) -> Result<(), EnqueueImageGenRequestError> {
+  Ok(())
+}
+
+#[derive(Debug)]
+pub enum DownloadMediaFileError {
+  IoError(std::io::Error),
+  MediaFileNotFound,
+  Other(anyhow::Error),
+}
+
+impl From<anyhow::Error> for DownloadMediaFileError {
+  fn from(err: anyhow::Error) -> Self {
+    DownloadMediaFileError::Other(err)
+  }
+}
+
+async fn query_and_download_media_file(
+  media_file_token: &MediaFileToken,
+  download_dir: &TempDir,
+  bucket_client: &BucketClient,
+  mysql_pool: &MySqlPool,
+) -> Result<PathBuf, DownloadMediaFileError> {
+  let media_file_result = get_media_file(
+    &media_file_token,
+    false,
+    mysql_pool
+  ).await;
+
+  let media_file= match media_file_result {
+    Ok(Some(result)) => result,
+    Ok(None) => {
+      warn!("could not find media file in database: {:?}", media_file_token);
+      return Err(DownloadMediaFileError::MediaFileNotFound)
+    }
+    Err(e) => {
+      error!("could not query media file: {:?}", e);
+      return Err(DownloadMediaFileError::Other(e))
+    }
+  };
+
+  let media_file_bucket_path = MediaFileBucketPath::from_object_hash(
+    &media_file.public_bucket_directory_hash,
+    media_file.maybe_public_bucket_prefix.as_deref(),
+    media_file.maybe_public_bucket_extension.as_deref());
+
+  let extension = media_file.maybe_public_bucket_extension.clone().unwrap_or_else(|| ".jpg".to_string());
+  let download_filename = format!("download_file_{}.{}", media_file.token.as_str(), extension);
+  let download_file_path = download_dir.path().join(download_filename);
+
+  info!("Downloading from bucket...");
+
+  bucket_client.download_file_to_disk(
+    &media_file_bucket_path.to_full_object_pathbuf(),
+    &download_file_path,
+  ).await?;
+
+  Ok(download_file_path)
+}

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/prompt/mod.rs

@@ -0,0 +1 @@
+pub mod enqueue_studio_image_generation_handler;

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/upload/mod.rs

@@ -0,0 +1 @@
+pub mod upload_snapshot_media_file_handler;

crates/service/web/storyteller_web/src/http_server/endpoints/image_studio/upload/upload_snapshot_media_file_handler.rs

@@ -0,0 +1,279 @@
+use std::collections::HashSet;
+use std::io::{BufReader, Bytes, Cursor, Read};
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use actix_multipart::form::tempfile::TempFile;
+use actix_multipart::form::text::Text;
+use actix_multipart::form::MultipartForm;
+use actix_multipart::Multipart;
+use actix_web::web::Json;
+use actix_web::{web, HttpRequest, HttpResponse};
+use log::{error, info, warn};
+use once_cell::sync::Lazy;
+use utoipa::ToSchema;
+
+use bucket_paths::legacy::typified_paths::public::media_files::bucket_file_path::MediaFileBucketPath;
+use enums::by_table::media_files::media_file_class::MediaFileClass;
+use enums::by_table::media_files::media_file_engine_category::MediaFileEngineCategory;
+use enums::by_table::media_files::media_file_type::MediaFileType;
+use enums::common::visibility::Visibility;
+use enums::no_table::style_transfer::style_transfer_name::StyleTransferName;
+use hashing::sha256::sha256_hash_bytes::sha256_hash_bytes;
+use http_server_common::request::get_request_ip::get_request_ip;
+use media::decode_basic_audio_info::decode_basic_audio_bytes_info;
+use mimetypes::mimetype_for_bytes::get_mimetype_for_bytes;
+use mimetypes::mimetype_to_extension::mimetype_to_extension;
+use mysql_queries::queries::idepotency_tokens::insert_idempotency_token::insert_idempotency_token;
+use mysql_queries::queries::media_files::create::specialized_insert::insert_media_file_from_file_upload::{insert_media_file_from_file_upload, InsertMediaFileFromUploadArgs, UploadType};
+use mysql_queries::queries::media_files::get::get_media_file::get_media_file;
+use tokens::tokens::media_files::MediaFileToken;
+
+use crate::http_server::endpoints::media_files::upload::upload_error::MediaFileUploadError;
+use crate::http_server::validations::validate_idempotency_token_format::validate_idempotency_token_format;
+use crate::state::server_state::ServerState;
+use crate::util::check_creator_tokens::{check_creator_tokens, CheckCreatorTokenArgs, CheckCreatorTokenResult};
+
+/// Form-multipart request fields.
+///
+/// IF VIEWING DOCS, PLEASE SEE BOTTOM OF PAGE `UploadSnapshotMediaFileForm` (Under "Schema") FOR DETAILS ON FIELDS AND NULLABILITY.
+#[derive(MultipartForm, ToSchema)]
+#[multipart(duplicate_field = "deny")]
+pub struct UploadSnapshotMediaFileForm {
+  /// UUID for request idempotency
+  #[multipart(limit = "2 KiB")]
+  #[schema(value_type = String, format = Binary)]
+  uuid_idempotency_token: Text<String>,
+
+  /// The uploaded "screenshot" / "snapshot" of the engine scene
+  #[multipart(limit = "512 MiB")]
+  #[schema(value_type = Vec<u8>, format = Binary)]
+  snapshot: TempFile,
+
+  /// Optional: If an engine scene was used to create this snapshot, provide it here to create a link
+  #[multipart(limit = "2 KiB")]
+  #[schema(value_type = Option<MediaFileToken>, format = Binary)]
+  scene_media_token: Option<Text<MediaFileToken>>,
+
+  /// Optional: Visibility of the scene
+  #[multipart(limit = "2 KiB")]
+  #[schema(value_type = Option<Visibility>, format = Binary)]
+  maybe_visibility: Option<Text<Visibility>>,
+}
+
+// Unlike the "upload" endpoints, which are pure inserts, these endpoints are *upserts*.
+#[derive(Serialize, ToSchema)]
+pub struct UploadSnapshotMediaFileSuccessResponse {
+  pub success: bool,
+
+  /// The media token of the "screenshot" / "snapshot" we took.
+  pub snapshot_media_token: MediaFileToken,
+}
+
+static ALLOWED_MIME_TYPES : Lazy<HashSet<&'static str>> = Lazy::new(|| {
+  HashSet::from([
+    "image/jpeg",
+    "image/png",
+    "image/gif",
+    "image/webp",
+  ])
+});
+
+/// Upload snapshots of the scenes created in the 3D image studio.
+#[utoipa::path(
+  post,
+  tag = "Image Studio",
+  path = "/v1/image_studio/scene_snapshot",
+  responses(
+    (status = 200, description = "Success Update", body = UploadSnapshotMediaFileSuccessResponse),
+    (status = 400, description = "Bad input", body = MediaFileUploadError),
+    (status = 401, description = "Not authorized", body = MediaFileUploadError),
+    (status = 429, description = "Too many requests", body = MediaFileUploadError),
+    (status = 500, description = "Server error", body = MediaFileUploadError),
+  ),
+  params(
+    (
+      "request" = UploadSnapshotMediaFileForm,
+      description = "IF VIEWING DOCS, PLEASE SEE BOTTOM OF PAGE `UploadSnapshotMediaFileForm` (Under 'Schema') FOR DETAILS ON FIELDS AND NULLABILITY."
+    ),
+  )
+)]
+pub async fn upload_snapshot_media_file_handler(
+  http_request: HttpRequest,
+  server_state: web::Data<Arc<ServerState>>,
+  MultipartForm(mut form): MultipartForm<UploadSnapshotMediaFileForm>,
+) -> Result<Json<UploadSnapshotMediaFileSuccessResponse>, MediaFileUploadError> {
+
+  let mut mysql_connection = server_state.mysql_pool
+      .acquire()
+      .await
+      .map_err(|err| {
+        error!("MySql pool error: {:?}", err);
+        MediaFileUploadError::ServerError
+      })?;
+
+  // ==================== READ SESSION ==================== //
+
+  let maybe_user_session = server_state
+      .session_checker
+      .maybe_get_user_session_from_connection(&http_request, &mut mysql_connection)
+      .await
+      .map_err(|e| {
+        error!("Session checker error: {:?}", e);
+        MediaFileUploadError::ServerError
+      })?;
+
+  let maybe_user_token = maybe_user_session
+      .as_ref()
+      .map(|session| session.get_strongly_typed_user_token());
+
+  let maybe_avt_token = server_state
+      .avt_cookie_manager
+      .get_avt_token_from_request(&http_request);
+
+  // ==================== BANNED USERS ==================== //
+
+  if let Some(ref user) = maybe_user_session {
+    if user.is_banned {
+      return Err(MediaFileUploadError::NotAuthorizedVerbose("user is banned".to_string()));
+    }
+  }
+
+  // ==================== HANDLE IDEMPOTENCY ==================== //
+
+  // TODO(bt, 2024-02-26): This should be a transaction.
+  let uuid_idempotency_token = form.uuid_idempotency_token.as_ref();
+
+  if let Err(reason) = validate_idempotency_token_format(uuid_idempotency_token) {
+    return Err(MediaFileUploadError::BadInput(reason));
+  }
+
+  insert_idempotency_token(uuid_idempotency_token, &mut *mysql_connection)
+      .await
+      .map_err(|err| {
+        error!("Error inserting idempotency token: {:?}", err);
+        MediaFileUploadError::BadInput("invalid idempotency token".to_string())
+      })?;
+
+  // ==================== UPLOAD METADATA ==================== //
+
+  let creator_set_visibility = form.maybe_visibility
+      .map(|visibility| visibility.0)
+      .or_else(|| {
+        maybe_user_session
+            .as_ref()
+            .map(|user_session| user_session.preferred_tts_result_visibility)
+      })
+      .unwrap_or(Visibility::default());
+
+  // ==================== USER DATA ==================== //
+
+  let ip_address = get_request_ip(&http_request);
+
+  // ==================== FILE VALIDATION ==================== //
+
+  let mut file_bytes = Vec::new();
+  form.snapshot.file.read_to_end(&mut file_bytes)
+      .map_err(|e| {
+        error!("Problem reading file: {:?}", e);
+        MediaFileUploadError::ServerError
+      })?;
+
+  let mimetype = get_mimetype_for_bytes(file_bytes.as_ref())
+      .map(|mimetype| mimetype.to_string())
+      .ok_or_else(|| {
+        warn!("Could not determine mimetype for file");
+        MediaFileUploadError::BadInput("Could not determine mimetype for file".to_string())
+      })?;
+
+  if !ALLOWED_MIME_TYPES.contains(mimetype.as_str()) {
+    // NB: Don't let our error message inject malicious strings
+    let filtered_mimetype = mimetype
+        .chars()
+        .filter(|c| c.is_ascii())
+        .filter(|c| c.is_alphanumeric() || *c == '/')
+        .collect::<String>();
+    return Err(MediaFileUploadError::BadInput(format!("unpermitted mime type: {}", &filtered_mimetype)));
+  }
+
+  // ==================== OTHER FILE METADATA ==================== //
+
+  let maybe_filename = form.snapshot.file_name.as_deref()
+      .as_deref()
+      .map(|filename| PathBuf::from(filename));
+
+  let extension = mimetype_to_extension(&mimetype)
+      .or_else(|| {
+        maybe_filename
+            .as_ref()
+            .and_then(|filename| filename.extension()) // TODO needs dot prefix
+            .and_then(|ext| ext.to_str())
+      })
+      .ok_or_else(|| {
+        warn!("Could not determine file extension for mimetype: {}", &mimetype);
+        MediaFileUploadError::ServerError
+      })?;
+
+  let extension = format!(".{extension}"); // NB: needs dot prefix
+
+  let file_size_bytes = file_bytes.len();
+
+  let hash = sha256_hash_bytes(&file_bytes)
+      .map_err(|io_error| {
+        error!("Problem hashing bytes: {:?}", io_error);
+        MediaFileUploadError::ServerError
+      })?;
+
+  // ==================== UPLOAD AND SAVE ==================== //
+
+  const PREFIX : Option<&str> = Some("snapshot_");
+
+  let public_upload_path = MediaFileBucketPath::generate_new(PREFIX, Some(&extension));
+
+  info!("Uploading media to bucket path: {}", public_upload_path.get_full_object_path_str());
+
+  server_state.public_bucket_client.upload_file_with_content_type(
+    public_upload_path.get_full_object_path_str(),
+    file_bytes.as_ref(),
+    &mimetype)
+      .await
+      .map_err(|e| {
+        warn!("Upload media bytes to bucket error: {:?}", e);
+        MediaFileUploadError::ServerError
+      })?;
+
+  let (token, record_id) = insert_media_file_from_file_upload(InsertMediaFileFromUploadArgs {
+    maybe_media_class: Some(MediaFileClass::Image),
+    media_file_type: MediaFileType::Image,
+    maybe_creator_user_token: maybe_user_token.as_ref(),
+    maybe_creator_anonymous_visitor_token: maybe_avt_token.as_ref(),
+    creator_ip_address: &ip_address,
+    creator_set_visibility,
+    upload_type: UploadType::Filesystem, // TODO(bt,2024-05-02): This should be a parameter and a well-known enum.
+    maybe_engine_category: None,
+    maybe_animation_type: None,
+    maybe_mime_type: Some(&mimetype),
+    file_size_bytes: file_size_bytes as u64,
+    maybe_duration_millis: None,
+    sha256_checksum: &hash,
+    maybe_title: Some("Snapshot"),
+    maybe_scene_source_media_file_token: None,
+    is_intermediate_system_file: true,
+    public_bucket_directory_hash: public_upload_path.get_object_hash(),
+    maybe_public_bucket_prefix: PREFIX,
+    maybe_public_bucket_extension: Some(&extension),
+    pool: &server_state.mysql_pool,
+  })
+      .await
+      .map_err(|err| {
+        warn!("New file creation DB error: {:?}", err);
+        MediaFileUploadError::ServerError
+      })?;
+
+  info!("new media file id: {} token: {:?}", record_id, &token);
+
+  Ok(Json(UploadSnapshotMediaFileSuccessResponse {
+    success: true,
+    snapshot_media_token: token,
+  }))
+}

crates/service/web/storyteller_web/src/http_server/endpoints/mod.rs

@@ -5,6 +5,7 @@ pub mod app_state;
 pub mod beta_keys;
 pub mod categories;
 pub mod comments;
+pub mod control_plane;
 pub mod conversion;
 pub mod download_job;
 pub mod engine;
@@ -12,6 +13,7 @@ pub mod events;
 pub mod featured_items;
 pub mod flags;
 pub mod image_gen;
+pub mod image_studio;
 pub mod inference_job;
 pub mod investor_demo;
 pub mod leaderboard;
@@ -30,9 +32,9 @@ pub mod subscriptions;
 pub mod tags;
 pub mod trending;
 pub mod tts;
-pub mod users;
 pub mod user_bookmarks;
 pub mod user_ratings;
+pub mod users;
 pub mod vocoders;
 pub mod voice_clone_requests;
 pub mod voice_conversion;

crates/service/web/storyteller_web/src/http_server/routes/add_control_plane_routes.rs

@@ -0,0 +1,24 @@
+use actix_http::body::MessageBody;
+use actix_service::ServiceFactory;
+use actix_web::dev::{ServiceRequest, ServiceResponse};
+use actix_web::{web, App, Error, HttpResponse};
+use crate::http_server::endpoints::control_plane::set_sora_secret_handler::set_sora_secret_handler;
+
+pub fn add_control_plane_routes<T, B> (app: App<T>) -> App<T>
+where
+    B: MessageBody,
+    T: ServiceFactory<
+      ServiceRequest,
+      Config = (),
+      Response = ServiceResponse<B>,
+      Error = Error,
+      InitError = (),
+    >,
+{
+  app.service(web::scope("/v1/control_plane")
+      .service(web::resource("/sora_secret")
+          .route(web::post().to(set_sora_secret_handler))
+          .route(web::head().to(|| HttpResponse::Ok()))
+      )
+  )
+}

crates/service/web/storyteller_web/src/http_server/routes/add_image_studio_routes.rs

@@ -0,0 +1,29 @@
+use crate::http_server::endpoints::image_studio::prompt::enqueue_studio_image_generation_handler::enqueue_studio_image_generation_handler;
+use crate::http_server::endpoints::image_studio::upload::upload_snapshot_media_file_handler::upload_snapshot_media_file_handler;
+use actix_http::body::MessageBody;
+use actix_service::ServiceFactory;
+use actix_web::dev::{ServiceRequest, ServiceResponse};
+use actix_web::{web, App, Error, HttpResponse};
+
+pub fn add_image_studio_routes<T, B> (app: App<T>) -> App<T>
+where
+  B: MessageBody,
+  T: ServiceFactory<
+    ServiceRequest,
+    Config = (),
+    Response = ServiceResponse<B>,
+    Error = Error,
+    InitError = (),
+  >,
+{
+  app.service(web::scope("/v1/image_studio")
+    .service(web::resource("/scene_snapshot")
+      .route(web::post().to(upload_snapshot_media_file_handler))
+      .route(web::head().to(|| HttpResponse::Ok()))
+    )
+    .service(web::resource("/prompt")
+      .route(web::post().to(enqueue_studio_image_generation_handler))
+      .route(web::head().to(|| HttpResponse::Ok()))
+    )
+  )
+}

crates/service/web/storyteller_web/src/http_server/routes/add_routes.rs

@@ -123,6 +123,8 @@ use crate::http_server::endpoints::w2l::get_w2l_template_use_count::get_w2l_temp
 use crate::http_server::endpoints::w2l::get_w2l_upload_template_job_status::get_w2l_upload_template_job_status_handler;
 use crate::http_server::endpoints::w2l::list_w2l_templates::list_w2l_templates_handler;
 use crate::http_server::endpoints::w2l::set_w2l_template_mod_approval::set_w2l_template_mod_approval_handler;
+use crate::http_server::routes::add_control_plane_routes::add_control_plane_routes;
+use crate::http_server::routes::add_image_studio_routes::add_image_studio_routes;
 use crate::http_server::routes::add_studio_gen2_routes::add_studio_gen2_routes;
 use crate::http_server::routes::beta_key_routes::add_beta_key_routes;
 use crate::http_server::routes::job_routes::add_job_routes;
@@ -171,8 +173,10 @@ pub fn add_routes<T, B> (app: App<T>, server_environment: ServerEnvironment) ->
   app = add_model_download_routes(app);
   app = add_workflow_routes(app);
   app = add_studio_gen2_routes(app);
+  app = add_image_studio_routes(app);
   app = add_job_routes(app);
   app = add_engine_routes(app); /* /v1/engine/... */
+  app = add_control_plane_routes(app); /* /v1/control_plane/... */
 
   // app = add_image_gen_routes(app);
 

crates/service/web/storyteller_web/src/http_server/routes/mod.rs

@@ -8,4 +8,6 @@ pub mod moderation_routes;
 pub mod user_routes;
 pub mod weights_routes;
 pub mod workflow_routes;
-mod tag_routes;
+mod tag_routes;
+mod add_image_studio_routes;
+mod add_control_plane_routes;