Some bug fixes and improvements for FK

There's still some glitches around selection and deletion.

.gitignore

@@ -66,3 +66,7 @@ scraps/
 # These should be built by the platform
 windows/sqlite/sqlite3.exp
 windows/sqlite/sqlite3.lib
+
+# ArtCraft Windows DLLs
+crates/desktop/src-tauri/*.dll
+windows_dlls/

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

@@ -0,0 +1,19 @@
+//! 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: &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
+    .with_intra_threads(1)? // Configure the number of threads used for inference
+    .with_execution_providers([
+      // Configure execution providers (e.g., CUDA, CoreML, CPU)
+      ExecutionProvider::CUDA(Default::default()),
+      ExecutionProvider::CoreML(Default::default()),
+      ExecutionProvider::CPU(Default::default()),
+    ])?
+    .with_model_from_file(onnx_model_file)?;
+  Ok(session)
+}

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

@@ -0,0 +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;

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

@@ -1,15 +1,14 @@
-use crate::ml::models::unet_model::UNetModel;
 use anyhow::anyhow;
+use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
 use candle_transformers::models::stable_diffusion::vae::AutoEncoderKL;
 use std::sync::{Arc, RwLock};
-
 // TODO: This data structure is gross. Generalize the locking and loading semantics for reuse.
 
 // Simple registry for now. We can build complex machinery that aids in 
 // VRAM utilization, disk space saving, intelligent scheduling, etc. in the future.
 pub struct ModelCache {
   vae: RwLock<Option<Arc<AutoEncoderKL>>>,
-  unet2: RwLock<Option<Arc<UNetModel>>>,
+  unet2: RwLock<Option<Arc<UNet2DConditionModel>>>,
 }
 
 impl ModelCache {
@@ -39,7 +38,7 @@ impl ModelCache {
     }
   }
 
-  pub fn set_unet(&self, vae: Arc<UNetModel>) -> anyhow::Result<()> {
+  pub fn set_unet(&self, vae: Arc<UNet2DConditionModel>) -> anyhow::Result<()> {
     match self.unet2.write() {
       Err(err) => Err(anyhow!("{}", err)),
       Ok(mut lock) => {
@@ -49,7 +48,7 @@ impl ModelCache {
     }
   }
 
-  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNetModel>>> {
+  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNet2DConditionModel>>> {
     match self.unet2.read() {
       Err(err) => Err(anyhow!("{}", err)),
       Ok(ref lock) => {

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

@@ -3,19 +3,51 @@ extern crate accelerate_src;
 #[cfg(feature = "mkl")]
 extern crate intel_mkl_src;
 
-use std::path::PathBuf;
 use candle_transformers::models::stable_diffusion;
+use std::path::{Path, PathBuf};
 
 use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
+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;
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppWeightsDir;
+
+
+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<_>>>()?;
@@ -50,20 +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_file = if first {
-  //  info!("ModelFile::Tokenizer");
-  //  ModelFile::Tokenizer
-  //} else {
-  //  info!("ModelFile::Tokenizer2");
-  //  ModelFile::Tokenizer2
-  //};
-
-  //let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
-
-  let tokenizer = weights_dir.model_path(&ModelType::ClipJson);
+  let tokenizer = clip_json_path;
 
   info!("Loading Clip Tokenizer path: {:?}", tokenizer);
 
@@ -85,14 +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.");
-  let clip_weights_file = if first { ModelFile::Clip } else { ModelFile::Clip2 };
+  
+  // TODO(bt,2025-03-23): These used to be two different clip models. We should only need one.
   let clip_weights = if first {
-    //clip_weights_file.get(clip_weights, sd_version, use_f16)?
-    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder)
+    println!(">>>>> CLIP 1");
+    clip_weights_path
   } else {
-    //clip_weights_file.get(clip2_weights, sd_version, use_f16)?
-    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder2)
+    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

@@ -1,72 +1,100 @@
-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::model_type::ModelType;
-use crate::ml::models::unet_model::UNetModel;
-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::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};
 
-pub struct Args<'a> {
+const DEFAULT_STRENGTH : f64 = 75.0;
+
+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<u8>,
-  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;
+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 dtype = *dtype;
+  let sd_version = *sd_version;
 
-  let weights_dir = app_data_root.weights_dir();
+  let img2img_strength = img2img_strength
+    .map(|s| remap_lcm_strength_range(s))
+    .unwrap_or(DEFAULT_STRENGTH);
+  
+  let img2img_strength = img2img_strength / 100.0;
 
-  let img2img_strength = args.i2i_strength.unwrap_or(75) as f64 / 100.0;
   let use_f16 = true;
 
-  //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);
-
   // 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);
 
-  //info!("Using seed: {}", 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"
@@ -76,72 +104,50 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
   // let use_guide_scale = guidance_scale > 1.0;
   let use_guide_scale = false;
 
-  //info!("Using guide scale: {}", use_guide_scale);
-
-  //info!("Checking if prompt is cached");
   let maybe_cached = prompt_cache.get_copy(&prompt)?;
 
   let mut text_embeddings = if let Some(tensor) = maybe_cached {
     tensor
   } else {
-    info!("Loading clip weights...");
-    //println!("Loading clip weights for {:?} with model {:?}", configs.sd_version, configs.sd_version.repo());
-    //let api = configs.hf_api.clone();
-
-    //let clip_path = api.model(configs.sd_version.repo().to_string()).get(if use_f16 { "text_encoder/model.fp16.safetensors" } else { "text_encoder/model.safetensors" })?;
-
-    //println!(">>>> CLIP REPO = {:?}", configs.sd_version.repo());
-    //println!(">>>> CLIP PATH = {:?}", clip_path);
-    //println!(">>>> USE_FP16 = {:?}", use_f16);
-    //let clip_weights = Some(clip_path.to_string_lossy().to_string());
-
-    let clip_weights = weights_dir.model_path(&ModelType::LykonDreamshaper7TextEncoderFp16);
-    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 repo = configs.sd_version.repo();
-      //println!(">>>>> VAE REPO = {:?}", repo);
-      //println!("Building VAE model from : {:?} ... (3)", repo);
-      //let vae_file = configs.hf_api.model(repo.to_string()).get("vae/diffusion_pytorch_model.safetensors")?;
-
-      let vae_file = weights_dir.model_path(&ModelType::LykonDreamshaper7Vae);
-
-      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);
 
@@ -156,11 +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.model_path(&ModelType::SimianLuoLcmDreamshaperV7Unet);
-
-      let in_channels = match configs.sd_version {
+      let in_channels = match sd_version {
         StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
         _ => 4,
       };
@@ -195,12 +199,10 @@ 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)?;
 
-      let unet = UNetModel::new_with_inner(unet);
-
       let unet = Arc::new(unet);
 
       model_cache.set_unet(unet.clone())?;
@@ -211,50 +213,40 @@ pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
   let init_latent_dist: DiagonalGaussianDistribution = vae.encode(&input_image)?;
 
-  // Parse and calculate the img2img strength
-  let img2img_strength = match i2i_strength {
-    None => 0.75f64,
-    Some(strength) => (strength as f64) / 100.0f64,
-  };
-
   // 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)?;
 
-  println!("Latents initialized successfully");
-  println!("Initial noise shape: {:?}", latents.shape());
-
-  println!("Starting diffusion process...");
-
   let mut latents = latents;
   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() };
 
     let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
 
     // Use the guidance scale embedding in the UNet inference with the proper method name
-    let mut noise_pred = match unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, &guidance_scale_embedding) {
+    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));
       },
     };
@@ -266,24 +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/mod.rs

@@ -1,4 +1,5 @@
 pub mod get_vae_scale;
 pub mod infer_clip_text_embeddings;
 pub mod stable_diffusion_pipeline;
-pub mod lcm_pipeline;
+pub mod lcm_pipeline;
+mod remap_lcm_strength_range;

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

@@ -0,0 +1,54 @@
+use primitives::numerics::remap_range_f64::{remap_range_f64, Args};
+
+/// Our LCM implementation uses strength [13, 100]. Strength below 13 will take 8 steps, 
+/// so we need to limit ourselves to this range from an input range of [0, 100].
+pub fn remap_lcm_strength_range(value: f64) -> f64 {
+  remap_range_f64(Args {
+    old_min: 0.0,
+    old_max: 100.0,
+    new_min: 13.0,
+    new_max: 100.0,
+    value,
+  })
+}
+
+#[cfg(test)]
+mod test {
+  use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
+  use assertor::{assert_that, FloatAssertion};
+
+  #[test]
+  fn old_max() {
+    assert_eq!(remap_lcm_strength_range(100.0), 100.0);
+  }
+  
+  #[test]
+  fn old_min() {
+    assert_eq!(remap_lcm_strength_range(0.0), 13.0);
+  }
+
+  #[test]
+  fn under_bounds() {
+    assert_eq!(remap_lcm_strength_range(-100.0), 13.0);
+  }
+  
+  #[test]
+  fn over_bounds() {
+    assert_eq!(remap_lcm_strength_range(200.0), 100.0);
+  }
+
+  #[test]
+  fn other_values() {
+    assert_that!(remap_lcm_strength_range(0.0)).with_abs_tol(0.001).is_approx_equal_to(13.000);
+    assert_that!(remap_lcm_strength_range(5.0)).with_abs_tol(0.001).is_approx_equal_to(17.350);
+    assert_that!(remap_lcm_strength_range(11.0)).with_abs_tol(0.001).is_approx_equal_to(22.570);
+    assert_that!(remap_lcm_strength_range(13.0)).with_abs_tol(0.001).is_approx_equal_to(24.310);
+    assert_that!(remap_lcm_strength_range(14.0)).with_abs_tol(0.001).is_approx_equal_to(25.180);
+    assert_that!(remap_lcm_strength_range(20.0)).with_abs_tol(0.001).is_approx_equal_to(30.400);
+    assert_that!(remap_lcm_strength_range(25.0)).with_abs_tol(0.001).is_approx_equal_to(34.750);
+    assert_that!(remap_lcm_strength_range(50.0)).with_abs_tol(0.001).is_approx_equal_to(56.500);
+    assert_that!(remap_lcm_strength_range(75.0)).with_abs_tol(0.001).is_approx_equal_to(78.250);
+    assert_that!(remap_lcm_strength_range(99.0)).with_abs_tol(0.001).is_approx_equal_to(99.130);
+    assert_that!(remap_lcm_strength_range(100.0)).with_abs_tol(0.001).is_approx_equal_to(100.000);
+  }
+}

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

@@ -1,76 +1,78 @@
+use std::path::Path;
 use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
 
-use crate::events::notification_event::{NotificationModelType, NotificationEvent};
 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::model_type::ModelType;
-use crate::ml::models::unet_model::UNetModel;
 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::state::app_config::AppConfig;
+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};
-use crate::state::app_dir::AppDataRoot;
 
-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,
@@ -91,20 +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
     };
@@ -113,14 +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()?;
     
@@ -129,33 +129,14 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No vae found in cache; loading...");
             
-            let mut notify_download_complete = false;
-            //if !vae_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
+            let vae_file = sdxl_turbo_vae_path.as_ref();
             
-            let vae_file = weights_dir.model_path(&ModelType::SdxlTurboVae);
-            
-            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);
             
             model_cache.set_vae(vae.clone())?;
-            
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-vae",
-                    model_type: NotificationModelType::Vae,
-                })?;
-            }
-            
+
             vae
         }
     };
@@ -167,32 +148,15 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
         None => {
             info!("No unet found in cache; loading...");
 
-            let mut notify_download_complete = false;
-            //if !unet_file.exists() {
-            if true {
-                notify_download_complete = true;
-                app.emit("notification", NotificationEvent::ModelDownloadStarted {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-            
-            let unet_file = weights_dir.model_path(&ModelType::SdxlTurboUnet);
+            let unet_weights_file = sdxl_turbo_unet_path.as_ref();
 
-            let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, 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);
 
             model_cache.set_unet(unet.clone())?;
 
-            if notify_download_complete {
-                app.emit("notification", NotificationEvent::ModelDownloadComplete {
-                    model_name: "sdxl-turbo-unet",
-                    model_type: NotificationModelType::Unet,
-                })?;
-            }
-
             unet
         }
     };
@@ -216,9 +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
     };
 
@@ -257,7 +221,7 @@ pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
 
         let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep)?;
 
-        let mut noise_pred = match unet.inference(&latent_model_input, timestep as f64, &text_embeddings) {
+        let mut noise_pred = match unet.forward(&latent_model_input, timestep as f64, &text_embeddings) {
             Ok(pred) => pred,
             Err(e) => {
                 println!("UNet inference failed with error: {}", e);

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

@@ -0,0 +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

@@ -0,0 +1,47 @@
+
+// TODO(bt,2025-03-13): When it's time for dynamic, non-const allocation, refer to
+//  http::header::name::HeaderName: https://github.com/hyperium/http/blob/master/src/header/name.rs
+//  as a pattern that will support both compile-time constants and dynamically allocated String-based variants.
+
+#[derive(Clone, Copy)]
+pub enum WeightFunction {
+  ImageGeneration,
+  ImageSegmentation,
+  TextEncoder,
+  TextTokenizer,
+  Unet,
+  Vae,
+}
+
+pub struct WeightDescriptor {
+  /// Name of the weight, eg. "Lykon Dreamshaper 7"
+  pub name: &'static str,
+
+  /// Name of the file on the filesystem, eg. lykon_dreamshaper_7_text_encoder.fp16.safetensors
+  pub filename: &'static str,
+
+  /// URL of the file in the R2 bucket,
+  /// eg. https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/lykon_dreamshaper_7_text_encoder.fp16.safetensors
+  pub r2_download_url: &'static str,
+
+  pub function: WeightFunction,
+}
+
+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",
+      WeightFunction::Unet => "UNET",
+      WeightFunction::Vae => "VAE",
+    }
+  }
+}
+
+impl WeightDescriptor {
+  pub fn to_descriptive_name(&self) -> String {
+    format!("{} ({})", self.name, self.function.as_descriptive_str())
+  }
+}

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

@@ -0,0 +1,17 @@
+pub (super) const R2_BUCKET_URL : &str = "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/";
+
+// TODO(bt,2025-03-13): Try to use keyword args from the call site.
+// NB: We're using a macro because we can't use a builder pattern and then concat string literals
+//  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::weights_registry::weight_descriptor::WeightDescriptor {
+        name: $name,
+        filename: $file_name,
+        r2_download_url: const_format::concatcp!(crate::weights_registry::weight_descriptor_builder::R2_BUCKET_URL, $file_name),
+        function: $weight_function,
+      }
+    };
+}
+
+pub(crate) use weight;

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

@@ -0,0 +1,65 @@
+use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
+use crate::weights_registry::weight_descriptor_builder::weight;
+
+pub const CLIP_JSON : WeightDescriptor = weight!(
+  "CLIP",
+  "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
+  WeightFunction::TextTokenizer
+);
+
+/// For LCM-based image generation
+pub const LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16 : WeightDescriptor = weight!(
+  "Lykon Dreamshaper 7",
+  "lykon_dreamshaper_7_text_encoder.fp16.safetensors",
+  WeightFunction::TextEncoder
+);
+
+/// For LCM-based image generation
+pub const LYKON_DEAMSHAPER_7_VAE : WeightDescriptor = weight!(
+  "Lykon Dreamshaper 7",
+  "lykon_dreamshaper_7_vae.safetensors",
+  WeightFunction::Vae
+);
+
+/// For LCM-based image generation
+pub const SIMIANLUO_LCM_DREAMSHAPER_V7_UNET : WeightDescriptor = weight!(
+  "SimianLuo LCM Dreamshaper v7",
+  "simianluo_lcm_dreamshaper_v7_unet.safetensors",
+  WeightFunction::Unet
+);
+
+/// Image segmentation and background removal
+/// dis_medium.onnx
+pub const DIS_MEDIUM_ONNX : WeightDescriptor = weight!(
+  "Dichotomous Image Segmentation Medium",
+  "dis_medium.onnx",
+  WeightFunction::ImageSegmentation
+);
+
+#[deprecated]
+pub const SDXL_TURBO_CLIP_TEXT_ENCODER : WeightDescriptor = weight!(
+  "SDXL Turbo CLIP",
+  "clip_text_encoder.safetensors",
+  WeightFunction::TextEncoder
+);
+
+#[deprecated]
+pub const SDXL_TURBO_CLIP_TEXT_ENCODER_2 : WeightDescriptor = weight!(
+  "SDXL Turbo CLIP #2",
+  "clip_text_encoder_2.safetensors",
+  WeightFunction::TextEncoder
+);
+
+#[deprecated]
+pub const SDXL_TURBO_VAE : WeightDescriptor = weight!(
+  "SDXL Turbo",
+  "diffusion_pytorch_model.vae.safetensors",
+  WeightFunction::Vae
+);
+
+#[deprecated]
+pub const SDXL_TURBO_UNET : WeightDescriptor = weight!(
+  "SDXL Turbo",
+  "diffusion_pytorch_model.unet.safetensors",
+  WeightFunction::Unet
+);

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
 
@@ -20,29 +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"] }
-dirs = "6.0.0"
+chrono = "0.4.40"
+const_format = "0.2.34"
 directories = "6.0.0"
+dirs = "6.0.0"
 futures = "0.3"
-hf-hub = "0.3.2"
+half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
 image = "0.25.5"
 log = "0.4"
 once_cell = "1.20.3"
+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"
@@ -51,25 +58,22 @@ 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"] }
-half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
 
-[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"
-]
+# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
+
+[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/download_models.rs

@@ -1,16 +1,12 @@
-use crate::ml::downloads::download_all_models::download_all_models;
 use crate::state::app_dir::AppDataRoot;
 use tauri::State;
 
 #[tauri::command]
 pub async fn download_models(
-  app_data_root: State<'_, AppDataRoot>,
+  _app_data_root: State<'_, AppDataRoot>,
 ) -> Result<String, String> {
-  
-  download_all_models(&app_data_root).await
-    .map_err(|err| {
-      "error downloading models".to_string()
-    })?;
-  
+
+  // NB: This endpoint has been removed for now as we now download models in the app lifecycle.
+
   Ok(("downloaded".to_string()))
 }

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

@@ -1,146 +0,0 @@
-use crate::ml::downloads::download_all_models::download_all_models;
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
-use crate::state::app_config::AppConfig;
-use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine};
-use bytes::BytesMut;
-use image::imageops::FilterType;
-use image::{DynamicImage, ImageFormat, ImageReader};
-use log::{error, info};
-use std::io::Cursor;
-use std::path::PathBuf;
-use tauri::{AppHandle, State};
-use crate::state::app_dir::AppDataRoot;
-
-const PROMPT_FILENAME : &str = "prompt.txt";
-const PROMPT: &str = "green goblin, blood dripping from lips, detailed, photorealistic, 8k";
-
-const NEGATIVE_PROMPT: &str = "bad quality, bad faces, poor quality, blurry faces, watermark";
-
-const RANDOM_SEED: u32 = 42;
-
-// TODO: Mark async https://github.com/tauri-apps/tauri/discussions/7737
-
-
-/// This handler takes an image (as a base64 encoded string) and a prompt and returns
-/// an image (as a base64-encoded string).
-#[tauri::command]
-pub async fn infer_image(
-  image: &str,
-  prompt: Option<String>,
-  strength: Option<u8>,
-  model_config: State<'_, AppConfig>,
-  model_cache: State<'_, ModelCache>,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
-) -> Result<String, String> {
-
-  let prompt = get_prompt_or_fallback(prompt);
-
-  info!("Prompt: {}", prompt);
-  info!("Strength: {:?}", strength);
-
-  let image = hydrate_base64_image(image)
-    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
-
-  // TODO(bt): Move this to another endpoint
-  download_all_models(&app_data_root).await
-    .map_err(|err| {
-      error!("couldn't download models: {:?}", err);
-      "Couldn't download models".to_string()
-    })?;
-
-  let result = do_infer_image(
-    &prompt, 
-    image, 
-    strength, 
-    &model_config, 
-    &model_cache, 
-    prompt_cache, 
-    app, 
-    &app_data_root
-  ).await;
-  
-  if let Err(err) = result.as_deref() {
-    error!("There was an error: {:?}", err);
-  }
-  
-  result
-}
-
-async fn do_infer_image(
-  prompt: &str,
-  image: DynamicImage,
-  strength: Option<u8>,
-  config: &AppConfig,
-  model_cache: &ModelCache,
-  prompt_cache: State<'_, PromptCache>,
-  app: AppHandle,
-  app_data_root: &AppDataRoot,
-) -> Result<String, String> {
-
-  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,
-    cfg_scale: config.cfg_scale,
-    app: &app,
-    app_data_root,
-    use_flash_attn: true,
-  };
-
-  match lcm_pipeline(args) {
-    Ok(image) => {
-      let mut bytes = Vec::with_capacity(1024*1024);
-      
-      image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)
-        .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-      let bytes = BASE64_STANDARD.encode(bytes);
-
-      Ok(bytes)
-    },
-    Err(e) => Err(format!("Failed to generate image: {}", e)),
-  }
-}
-
-fn get_prompt_or_fallback(user_prompt: Option<String>) -> String {
-  let user_prompt = user_prompt.map(|prompt| prompt.trim().to_string())
-    .filter(|prompt| !prompt.is_empty());
-
-  if let Some(prompt) = user_prompt {
-    return prompt;
-  }
-
-  let prompt_file = PathBuf::from(PROMPT_FILENAME)
-    .canonicalize()
-    .unwrap_or_else(|_| PathBuf::from(PROMPT_FILENAME));
-
-  std::fs::read_to_string(&prompt_file)
-    .map_err(|err| format!("Failed to read prompt file: {}", err))
-    .unwrap_or_else(|_| {
-      error!("Failed to read prompt file: {:?}", prompt_file);
-      PROMPT.to_string()
-    })
-    .trim()
-    .to_string()
-}
-
-fn hydrate_base64_image(base64_image: &str) -> anyhow::Result<DynamicImage> {
-  let bytes = BASE64_STANDARD.decode(base64_image)?;
-
-  let image = ImageReader::new(Cursor::new(bytes))
-    .with_guessed_format()?
-    .decode()?;
-
-  // TODO(bt,2025-02-17): Running out of vram with full image buffer size
-  let image = image.resize(512, 512, FilterType::CatmullRom);
-
-  Ok(image)
-}

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

@@ -1,4 +1,7 @@
 pub mod download_models;
 pub mod flip_image;
-pub mod image_endpoint;
-pub mod test_counter;
+pub mod inpaint_image_endpoint;
+pub mod realtime_image_endpoint;
+pub mod remove_background_endpoint;
+pub mod save_image_endpoint;
+pub mod text_to_image_endpoint;

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

@@ -0,0 +1,148 @@
+#[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";
+
+const NEGATIVE_PROMPT: &str = "bad quality, bad faces, poor quality, blurry faces, watermark";
+
+const RANDOM_SEED: u32 = 42;
+
+
+/// This handler takes an image (as a base64 encoded string) and a prompt and returns
+/// an image (as a base64-encoded string).
+#[tauri::command]
+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>,
+  app_data_root: State<'_, AppDataRoot>,
+  app: AppHandle,
+) -> Result<String, String> {
+  info!("infer_image endpoint called.");
+
+  let prompt = get_prompt_or_fallback(prompt);
+
+  info!("Strength: {:?}; Prompt: {}", strength, prompt);
+
+  let image = decode_base64_image(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 mut image = image.resize(512, 512, FilterType::CatmullRom);
+
+  let seed = seed.or(model_config.seed);
+
+  #[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...");
+
+  Ok(bytes)
+}
+
+#[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_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,
+    prompt_cache: &prompt_cache,
+    img2img_strength: strength,
+    cfg_scale: config.cfg_scale,
+    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)
+    .map_err(|err| format!("failure to encode image: {:?}", err))?;
+
+  Ok(image)
+}
+
+fn get_prompt_or_fallback(user_prompt: Option<String>) -> String {
+  let user_prompt = user_prompt.map(|prompt| prompt.trim().to_string())
+    .filter(|prompt| !prompt.is_empty());
+
+  if let Some(prompt) = user_prompt {
+    return prompt;
+  }
+
+  let prompt_file = PathBuf::from(PROMPT_FILENAME)
+    .canonicalize()
+    .unwrap_or_else(|_| PathBuf::from(PROMPT_FILENAME));
+
+  std::fs::read_to_string(&prompt_file)
+    .map_err(|err| format!("Failed to read prompt file: {}", err))
+    .unwrap_or_else(|_| {
+      error!("Failed to read prompt file: {:?}", prompt_file);
+      PROMPT.to_string()
+    })
+    .trim()
+    .to_string()
+}

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

@@ -0,0 +1,45 @@
+//! Adapted from https://github.com/dnanhkhoa/rust-background-removal
+
+#[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::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,
+  app_data_root: State<'_, AppDataRoot>,
+) -> Result<String, String> {
+  info!("remove_background endpoint called.");
+
+  let mut image = decode_base64_image(image)
+      .map_err(|err| format!("Couldn't hydrate image from base64: {}", 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))?;
+
+  Ok(image)
+}
+
+#[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);
+
+  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

@@ -0,0 +1,114 @@
+#[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, 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 tauri::{AppHandle, State};
+
+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.
+static IMAGE : Lazy<DynamicImage> = Lazy::new(||{
+  ImageReader::new(Cursor::new(PNG_BYTES))
+    .with_guessed_format()
+    .expect("failed to detect png format")
+    .decode()
+    .expect("failed to decode image")
+});
+
+#[tauri::command]
+pub async fn text_to_image(
+  prompt: String,
+  model_config: State<'_, AppConfig>,
+  model_cache: State<'_, ModelCache>,
+  prompt_cache: State<'_, PromptCache>,
+  app: AppHandle,
+  app_data_root: State<'_, AppDataRoot>,
+) -> Result<String, String> {
+  info!("text_to_image endpoint called.");
+
+  let mut image = IMAGE.resize(512, 512, FilterType::CatmullRom);
+
+  #[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;
+
+    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...");
+
+  Ok(bytes)
+}
+
+#[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,
+) -> Result<RgbImage, String> {
+  
+  let args = FluxPipelineArgs {
+    image: &image,
+    prompt: prompt.to_string(),
+    model_cache,
+    prompt_cache: &prompt_cache,
+    app: &app,
+    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 = flux_pipeline(args)
+    .map_err(|err| format!("failure to encode image: {:?}", err))?;
+
+  Ok(image)
+}

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

@@ -1,13 +1,16 @@
 use serde::Serialize;
 
-
 #[derive(Clone, Copy, Serialize)]
 #[serde(rename_all = "snake_case")]
 pub enum NotificationModelType {
+  ImageSegmentation,
+  TextEncoder,
+  TextTokenizer,
   Unet,
   Vae,
   Json,
   ClipEncoder,
+  FluxModel,
 }
 
 #[derive(Clone, Serialize)]
@@ -26,5 +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

@@ -1,17 +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::image_endpoint::infer_image;
-use crate::endpoints::test_counter::test_counter;
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
+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::save_image_endpoint::save_image;
+use crate::endpoints::text_to_image_endpoint::text_to_image;
 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 tauri_plugin_log::Target;
 use tauri_plugin_log::TargetKind;
 
@@ -39,9 +45,15 @@ 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
+      // log_environment_details();
+
       //if cfg!(debug_assertions) {
       //  app.handle().plugin(
       //    tauri_plugin_log::Builder::default()
@@ -60,9 +72,14 @@ pub fn run() {
     .manage(model_cache)
     .manage(app_data_root)
     .invoke_handler(tauri::generate_handler![
-      test_counter, 
-      infer_image,
       download_models,
+      flip_image,
+      infer_image,
+      inpaint_image,
+      remove_background,
+      save_image,
+      text_to_image,
+      text_to_image,
     ])
     .run(tauri::generate_context!())
     .expect("error while running tauri application");

crates/desktop/src-tauri/src/ml/downloads/download_all_models.rs

@@ -1,15 +0,0 @@
-use crate::ml::downloads::download_model_registry_file::download_model_registry_file;
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppDataRoot;
-
-pub async fn download_all_models(app_data_root: &AppDataRoot) -> anyhow::Result<()> {
-  let weights_dir = app_data_root.weights_dir();
-  
-  //download_model_registry_file(ModelType::ClipJson, weights_dir).await?;
-  //download_model_registry_file(ModelType::SdxlTurboUnet, weights_dir).await?;
-  //download_model_registry_file(ModelType::SdxlTurboVae, weights_dir).await?;
-  //download_model_registry_file(ModelType::SdxlTurboClipEncoder, weights_dir).await?;
-  //download_model_registry_file(ModelType::SdxlTurboClipEncoder2, weights_dir).await?;
-  
-  Ok(())
-}

crates/desktop/src-tauri/src/ml/downloads/download_model_registry_file.rs

@@ -1,30 +0,0 @@
-use crate::ml::model_type::ModelType;
-use crate::state::app_dir::AppWeightsDir;
-use log::info;
-use std::fs::File;
-use std::path::{Path, PathBuf};
-
-pub async fn download_model_registry_file(
-  model_type: ModelType,
-  weights_dir: &AppWeightsDir,
-) -> anyhow::Result<PathBuf> {
-  
-  let path = weights_dir.model_path(&model_type);
-
-  if path.exists() {
-    return Ok(path);
-  }
-
-  info!("downloading model: {:?} to {:?}", model_type, path);
-
-  let body = reqwest::get(model_type.get_download_url())
-    .await?
-    .bytes()
-    .await?;
-  
-  let _file = File::create(&path)?;
-  
-  std::fs::write(&path, body.as_ref())?;
-  
-  Ok(path)
-}

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

@@ -1,2 +0,0 @@
-pub mod download_model_registry_file;
-pub mod download_all_models;

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

@@ -1,8 +0,0 @@
-pub mod image;
-pub mod model_cache;
-pub mod model_file;
-pub mod models;
-pub mod prompt_cache;
-pub mod stable_diffusion;
-pub mod model_type;
-pub mod downloads;

crates/desktop/src-tauri/src/ml/model_type.rs

@@ -1,86 +0,0 @@
-use crate::events::notification_event::NotificationModelType;
-use crate::state::app_dir::AppWeightsDir;
-use std::path::PathBuf;
-
-#[derive(Copy, Clone, Debug)]
-pub enum ModelType {
-  ClipJson,
-  SdxlTurboUnet,
-  SdxlTurboVae,
-  SdxlTurboClipEncoder,
-  SdxlTurboClipEncoder2,
-  SimianLuoLcmDreamshaperV7Unet,
-  LykonDreamshaper7Vae,
-  LykonDreamshaper7TextEncoderFp16,
-}
-
-impl ModelType {
-  pub fn get_name(&self) -> &'static str {
-    match self {
-      Self::ClipJson => "CLIP (JSON)",
-      Self::SdxlTurboUnet => "SDXL Turbo UNET",
-      Self::SdxlTurboVae => "SDXL Turbo VAE",
-      Self::SdxlTurboClipEncoder => "Clip Encoder #1",
-      Self::SdxlTurboClipEncoder2 => "Clip Encoder #2",
-      Self::SimianLuoLcmDreamshaperV7Unet => "SimianLuo LCM Dreamshaper v7 UNET",
-      Self::LykonDreamshaper7Vae => "Lykon Dreamshaper 7 VAE",
-      Self::LykonDreamshaper7TextEncoderFp16 => "Lykon Dreamshaper Text Encoder (fp16)",
-    }
-  }
-
-  pub fn get_notification_type(&self) -> NotificationModelType {
-    match self {
-      Self::ClipJson => NotificationModelType::Json,
-      Self::SdxlTurboUnet => NotificationModelType::Unet,
-      Self::SdxlTurboVae => NotificationModelType::Vae,
-      Self::SdxlTurboClipEncoder => NotificationModelType::ClipEncoder,
-      Self::SdxlTurboClipEncoder2 => NotificationModelType::ClipEncoder,
-      Self::SimianLuoLcmDreamshaperV7Unet => NotificationModelType::Unet,
-      Self::LykonDreamshaper7Vae => NotificationModelType::Vae,
-      Self::LykonDreamshaper7TextEncoderFp16 => NotificationModelType::ClipEncoder,
-    }
-  }
-
-  pub fn get_hf_id(&self) -> Option<&'static str> {
-    match self {
-      Self::ClipJson => Some("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"),
-      Self::SdxlTurboUnet 
-      | Self::SdxlTurboVae
-      | Self::SdxlTurboClipEncoder
-      | Self::SdxlTurboClipEncoder2 => Some("stabilityai/sdxl-turbo"),
-      Self::SimianLuoLcmDreamshaperV7Unet => Some("SimianLuo/LCM_Dreamshaper_v7"),
-      Self::LykonDreamshaper7Vae
-      | Self::LykonDreamshaper7TextEncoderFp16 => Some("Lykon/dreamshaper-7"),
-    }
-  }
-
-  pub fn get_download_url(&self) -> &'static str {
-    match self {
-      Self::ClipJson => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/tokenizer.json",
-      Self::SdxlTurboUnet => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.unet.safetensors",
-      Self::SdxlTurboVae => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/diffusion_pytorch_model.vae.safetensors",
-      Self::SdxlTurboClipEncoder => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder.safetensors",
-      Self::SdxlTurboClipEncoder2 => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/clip_text_encoder_2.safetensors",
-      Self::SimianLuoLcmDreamshaperV7Unet => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/simianluo_lcm_dreamshaper_v7_unet.safetensors",
-      Self::LykonDreamshaper7Vae => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/lykon_dreamshaper_7_vae.safetensors",
-      Self::LykonDreamshaper7TextEncoderFp16 => "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/lykon_dreamshaper_7_text_encoder.fp16.safetensors",
-    }
-  }
-
-  pub fn get_filename(&self) -> &'static str {
-    match self {
-      Self::ClipJson => "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
-      Self::SdxlTurboUnet => "diffusion_pytorch_model.unet.safetensors",
-      Self::SdxlTurboVae => "diffusion_pytorch_model.vae.safetensors",
-      Self::SdxlTurboClipEncoder => "clip_text_encoder.safetensors",
-      Self::SdxlTurboClipEncoder2 => "clip_text_encoder_2.safetensors",
-      Self::SimianLuoLcmDreamshaperV7Unet => "simianluo_lcm_dreamshaper_v7_unet.safetensors",
-      Self::LykonDreamshaper7Vae => "lykon_dreamshaper_7_vae.safetensors",
-      Self::LykonDreamshaper7TextEncoderFp16 => "lykon_dreamshaper_7_text_encoder.fp16.safetensors",
-    }
-  }
-  
-  pub fn get_path(&self, weights_dir: &AppWeightsDir) -> PathBuf {
-    weights_dir.model_path(self)
-  }
-}

crates/desktop/src-tauri/src/ml/models/lazy_load_vae_model.rs

@@ -1,130 +0,0 @@
-use crate::ml::model_file::StableDiffusionVersion;
-use anyhow::anyhow;
-use candle_core::{DType, Device, Tensor};
-use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use hf_hub::api::sync::Api;
-use std::sync::{Arc, RwLock};
-
-pub struct LazyLoadVaeModel {
-  lazy_model: Arc<RwLock<Option<AutoEncoderKL>>>,
-  model_configs: VaeConfigs,
-}
-
-// TODO: Maybe we pass these via a central config read at startup, passed as an Arc<ConfigT>
-pub struct VaeConfigs {
-  sd_config: StableDiffusionConfig,
-  sd_version: StableDiffusionVersion,
-  device: Device,
-  dtype: DType,
-  hf_api: Api,
-}
-
-impl LazyLoadVaeModel {
-  pub fn new(
-    sd_config: &StableDiffusionConfig,
-    sd_version: &StableDiffusionVersion,
-    hf_api: &Api,
-    device: &Device,
-    dtype: DType,
-  ) -> anyhow::Result<Self> {
-    println!("building vae model... (1)");
-    
-    Ok(Self {
-      lazy_model: Arc::new(RwLock::new(None)),
-      model_configs: VaeConfigs {
-        sd_config: sd_config.clone(),
-        sd_version: sd_version.clone(),
-        device: device.clone(),
-        hf_api: hf_api.clone(),
-        dtype
-      }
-    })
-  }
-
-  pub fn encode(&self, xs: &Tensor) -> anyhow::Result<DiagonalGaussianDistribution> {
-    println!("vae encode...");
-    match self.try_encode(xs) {
-      Err(err) => Err(err),
-      Ok(Some(result)) => Ok(result),
-      Ok(None) => {
-        // Model wasn't previously loaded
-        self.load_model()?;
-        self.try_encode(xs)?
-          .ok_or(anyhow!("model was not previously loaded"))
-      }
-    }
-  }
-  
-  pub fn decode(&self, xs: &Tensor) -> anyhow::Result<Tensor> {
-    println!("vae decode...");
-    match self.try_decode(xs) {
-      Err(err) => Err(err),
-      Ok(Some(result)) => Ok(result),
-      Ok(None) => {
-        // Model wasn't previously loaded
-        self.load_model()?;
-        self.try_decode(xs)?
-          .ok_or(anyhow!("model was not previously loaded"))
-      }
-    }
-  }
-
-  fn try_encode(&self, xs: &Tensor) -> anyhow::Result<Option<DiagonalGaussianDistribution>> {
-    match self.lazy_model.read() {
-      Err(err) => Err(anyhow!("lock error: {:?}", err)),
-      Ok(model) => {
-        match &*model {
-          None => Ok(None), // Model is not yet loaded
-          Some(model) => {
-            Ok(Some(model.encode(xs)?))
-          }
-        }
-      }
-    }
-  }
-
-  fn try_decode(&self, xs: &Tensor) -> anyhow::Result<Option<Tensor>> {
-    match self.lazy_model.read() {
-      Err(err) => Err(anyhow!("lock error: {:?}", err)),
-      Ok(model) => {
-        match &*model {
-          None => Ok(None), // Model is not yet loaded
-          Some(model) => {
-            Ok(Some(model.decode(xs)?))
-          }
-        }
-      }
-    }
-  }
-
-  fn load_model(&self) -> anyhow::Result<()> {
-    println!("load vae");
-    match self.lazy_model.write() {
-      Err(err) => return Err(anyhow!("lock error: {:?}", err)),
-      Ok(mut maybe_model) => {
-        match &*maybe_model {
-          Some(_model) => {} // Already loaded; fall through
-          None => {
-            let repo = self.model_configs.sd_version.repo();
-
-            println!("Building VAE model from : {:?} ... (3)", repo);
-            
-            let vae_file = self.model_configs.hf_api.model(repo.to_string())
-              .get("vae/diffusion_pytorch_model.safetensors")?;
-
-            println!("Building VAE model from file {:?}...", &vae_file);
-            let vae = self.model_configs
-              .sd_config
-              .build_vae(vae_file, &self.model_configs.device, self.model_configs.dtype)?;
-
-            println!("Built VAE...");
-            *maybe_model = Some(vae);
-          }
-        }
-      }
-    }
-    Ok(())
-  }
- 
-}

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

@@ -1,2 +0,0 @@
-pub mod unet_model;
-pub mod lazy_load_vae_model;

crates/desktop/src-tauri/src/ml/models/unet_model.rs

@@ -1,30 +0,0 @@
-use candle_core::{DType, Device, Tensor};
-use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use std::path::Path;
-
-pub struct UNetModel {
-  model: UNet2DConditionModel,
-}
-
-impl UNetModel {
-  pub fn new<P: AsRef<Path>>(sd_config: &StableDiffusionConfig, unet_weights_file: P, device: &Device, dtype: DType) -> anyhow::Result<Self> {
-    println!("building unet model... (1)");
-
-    let model = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)?;
-
-    Ok(Self { model })
-  }
-
-  pub fn new_with_inner(model: UNet2DConditionModel) -> Self {
-    Self { model }
-  }
-
-  pub fn inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Tensor> {
-    Ok(self.model.forward(&latent_model_input, timestep, &text_embeddings)?)
-  }
-
-  pub fn forward_with_guidance(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor, guidance_scale_embedding: &Tensor) -> anyhow::Result<Tensor> {
-    Ok(self.model.forward_with_guidance(&latent_model_input, timestep, &text_embeddings, Some(guidance_scale_embedding))?)
-  }
-}

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

@@ -1,17 +1,22 @@
-use crate::ml::model_type::ModelType;
 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};
+
 const DEFAULT_DATA_DIR : &str = "artcraft";
 const ASSETS_SUBDIRECTORY : &str = "assets";
 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 {
@@ -19,6 +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)]
@@ -69,12 +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,
     })
   }
   
@@ -93,6 +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 {
@@ -116,6 +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 {
@@ -139,9 +173,9 @@ impl AppWeightsDir {
   pub fn path(&self) -> &Path {
     &self.path
   }
-  
-  pub fn model_path(&self, model_registry: &ModelType) -> PathBuf {
-    self.path.join(model_registry.get_filename())
+
+  pub fn weight_path(&self, descriptor: &WeightDescriptor) -> PathBuf {
+    self.path.join(descriptor.filename)
   }
 }
 

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

@@ -1,10 +1,11 @@
 use crate::events::notification_event::{NotificationEvent, NotificationModelType};
-use crate::ml::model_type::ModelType;
 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;
@@ -12,6 +13,7 @@ use std::sync::{Arc, Mutex};
 use tauri::{AppHandle, Emitter};
 use tempfile::NamedTempFile;
 use tokio::task::JoinHandle;
+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;
@@ -39,22 +41,32 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
 
   let mut download_queue = VecDeque::new();
 
-  // TODO: Automatic enqueue of known important models + enqueue new models on-demand
-  download_queue.push_back(ModelType::ClipJson);
-  //download_queue.push_back(ModelType::SdxlTurboUnet);
-  //download_queue.push_back(ModelType::SdxlTurboVae);
-  download_queue.push_back(ModelType::SdxlTurboClipEncoder); // TODO(bt): Why is this still needed?
-  //download_queue.push_back(ModelType::SdxlTurboClipEncoder2);
-  download_queue.push_back(ModelType::SimianLuoLcmDreamshaperV7Unet);
-  download_queue.push_back(ModelType::LykonDreamshaper7Vae);
-  download_queue.push_back(ModelType::LykonDreamshaper7TextEncoderFp16);
+  download_queue.push_back(CLIP_JSON);
+  download_queue.push_back(LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16);
+  download_queue.push_back(LYKON_DEAMSHAPER_7_VAE);
+  download_queue.push_back(SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
+  download_queue.push_back(SDXL_TURBO_CLIP_TEXT_ENCODER); // TODO(bt): Why is this still needed?
+  download_queue.push_back(DIS_MEDIUM_ONNX);
+  //download_queue.push_back(SDXL_TURBO_CLIP_TEXT_ENCODER_2); // TODO(bt): REMOVE
+  //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() {
-    let url = model.get_download_url().to_string();
+    let url = model.r2_download_url.to_string();
 
-    let filename = app_data_root.weights_dir().model_path(&model);
+    let filename = app_data_root.weights_dir().weight_path(&model);
 
     if filename.exists() {
       info!("File already exists: {:?}", &filename);
@@ -75,9 +87,18 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
       }
     };
 
+    let model_type = match model.function {
+      WeightFunction::TextEncoder => NotificationModelType::TextEncoder,
+      WeightFunction::TextTokenizer => NotificationModelType::TextTokenizer,
+      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 {
-      model_name: model.get_name(),
-      model_type: model.get_notification_type(),
+      model_name: model.name,
+      model_type,
     });
 
     if let Err(err) = result {
@@ -102,8 +123,8 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
         lock.push(ModelDownloadStatusChannel {
           receiver: rx,
           final_file_path: filename.clone(),
-          model_name: model.get_name().to_string(),
-          model_type: model.get_notification_type(),
+          model_name: model.to_descriptive_name(),
+          model_type,
         })
       }
     }

crates/desktop/src-tauri/src/utils/image/decode_base64_image.rs

@@ -0,0 +1,14 @@
+use base64::prelude::BASE64_STANDARD;
+use base64::Engine;
+use image::{DynamicImage, ImageReader};
+use std::io::Cursor;
+
+pub fn decode_base64_image(base64_image: &str) -> anyhow::Result<DynamicImage> {
+  let bytes = BASE64_STANDARD.decode(base64_image)?;
+
+  let image = ImageReader::new(Cursor::new(bytes))
+    .with_guessed_format()?
+    .decode()?;
+
+  Ok(image)
+}

crates/desktop/src-tauri/src/utils/image/encode_dynamic_image_base64_png.rs

@@ -0,0 +1,14 @@
+use base64::prelude::BASE64_STANDARD;
+use base64::Engine;
+use image::{DynamicImage, ImageFormat};
+use std::io::Cursor;
+
+pub fn encode_dynamic_image_base64_png(image: DynamicImage) -> anyhow::Result<String> {
+  let mut bytes = Vec::with_capacity(1024*1024);
+
+  image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)?;
+
+  let bytes = BASE64_STANDARD.encode(bytes);
+
+  Ok(bytes)
+}

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

@@ -0,0 +1,16 @@
+use base64::prelude::BASE64_STANDARD;
+use base64::Engine;
+use image::{ImageFormat, RgbImage};
+use std::io::Cursor;
+
+pub fn encode_rgb_image_base64_png(
+  image: RgbImage
+) -> anyhow::Result<String> {
+  let mut bytes = Vec::with_capacity(1024*1024);
+
+  image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)?;
+
+  let bytes = BASE64_STANDARD.encode(bytes);
+
+  Ok(bytes)
+}

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

@@ -0,0 +1,3 @@
+pub mod decode_base64_image;
+pub mod encode_dynamic_image_base64_png;
+pub mod encode_rgb_image_base64_png;

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

@@ -0,0 +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};
+
+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));
+}

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

@@ -0,0 +1,2 @@
+pub mod image;
+pub mod log_environment_details;

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/data/primitives/Cargo.toml

@@ -32,3 +32,4 @@ my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }
 
 [dev-dependencies]
 testing = { path = "../../testing" }
+assertor = "0.0.3"

crates/lib/data/primitives/src/numerics/mod.rs

@@ -1,5 +1,6 @@
 pub mod i32_to_u32_zero_clamped;
 pub mod i64_to_u64_zero_clamped;
+pub mod remap_range_f64;
 pub mod u64_to_i32_saturating;
 pub mod u64_to_i64_saturating;
 pub mod u64_to_u32_saturating;

crates/lib/data/primitives/src/numerics/remap_range_f64.rs

@@ -0,0 +1,131 @@
+
+pub struct Args {
+  pub old_min: f64,
+  pub old_max: f64,
+  pub new_min: f64,
+  pub new_max: f64,
+  
+  /// The value to remap to the new range.
+  pub value: f64,
+}
+
+pub fn remap_range_f64(args: Args) -> f64 {
+  let old_range = args.old_max - args.old_min;
+  let new_range = args.new_max -  args.new_min;
+
+  let new_value = (args.value -  args.old_min) * new_range / old_range + args.new_min;
+  
+  if new_value > args.new_max {
+    args.new_max
+  } else if new_value < args.new_min { 
+    args.new_min
+  } else {
+    new_value
+  }
+}
+
+#[cfg(test)]
+mod tests {
+  use crate::numerics::remap_range_f64::{remap_range_f64, Args};
+  use assertor::{assert_that, FloatAssertion};
+
+  #[test]
+  fn test_base_cases() {
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: 0.0,
+      new_max: 1.0,
+      value: 0.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
+    
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: 0.0,
+      new_max: 1.0,
+      value: 1.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(1.0);
+  }
+
+  #[test]
+  fn test_maxima() {
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 100.0,
+      new_min: 0.0,
+      new_max: 1.0,
+      value: 100.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(1.0);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: 0.0,
+      new_max: 50.0,
+      value: 1.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(50.0);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: -100.0,
+      new_max: 100.0,
+      value: 1.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(100.0);
+  }
+
+  #[test]
+  fn test_minima() {
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 100.0,
+      new_min: 0.0,
+      new_max: 1.0,
+      value: 0.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: 0.0,
+      new_max: 50.0,
+      value: 0.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: -100.0,
+      new_max: 100.0,
+      value: 0.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(-100.0);
+  }
+
+  #[test]
+  fn test_other_values() {
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 100.0,
+      new_min: 0.0,
+      new_max: 1.0,
+      value: 50.0,
+    })).with_abs_tol(0.01).is_approx_equal_to(0.50);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: 0.0,
+      new_max: 50.0,
+      value: 0.10,
+    })).with_abs_tol(0.01).is_approx_equal_to(5.0);
+
+    assert_that!(remap_range_f64(Args {
+      old_min: 0.0,
+      old_max: 1.0,
+      new_min: -100.0,
+      new_max: 100.0,
+      value: 0.25,
+    })).with_abs_tol(0.01).is_approx_equal_to(-50.0);
+  }
+}

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>>,
+}