don't download sdxl anymore

stop downloading clip tokenizer from hf
vae wasn't cached in memory
2026-10-25 19:29:29 -04:00 · 2025-03-13 00:31:51 -04:00 · 2025-03-13 00:24:06 -04:00 · 2025-03-12 23:58:52 -04:00 · 2025-03-12 23:54:42 -04:00 · 2025-03-12 23:38:49 -04:00
3596 changed files with 3387 additions and 176181 deletions
@@ -1,4 +1,4 @@
-hakari-package = "workspace-build-acceleration"
+hakari-package = "my-workspace-hack"

 resolver = "1"

@@ -11,4 +11,4 @@ platforms = [
    "x86_64-pc-windows-msvc",
 ]

-exact-versions = true
+exact-versions = true
@@ -32,9 +32,6 @@ target/
 crates/lib/database_queries/target/
 crates/lib/sqlite_queries/target/

-# Javascript build directories
-node_modules/
-
 # Vim noise
 *.swo
 *.swp
@@ -69,7 +66,3 @@ 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/
@@ -1,8 +1,6 @@
 [workspace]
 resolver = "2"

-# This block defines the workspace members.
-# When we add a library or binary crate, it needs an entry here.
 members = [
  # Types: common strongly-typed tokens, ids, payloads, etc.
  "crates/_types/composite_identifiers",
@@ -24,8 +22,7 @@ members = [
  "crates/cli/production/migration-tool",

  # Desktop apps
-  "crates/desktop/tauri-4o",
-  "crates/desktop/tauri-realtime",
+  "crates/desktop/src-tauri",

  # Shared libraries
  "crates/lib/actix_cors_configs",
@@ -33,7 +30,6 @@ 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",
@@ -60,20 +56,16 @@ 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/machine_learning/ml_models",
-  "crates/lib/machine_learning/ml_weights_registry",
-  "crates/lib/machine_learning/prompt_engineering",
  "crates/lib/markdown",
  "crates/lib/migration",
  "crates/lib/mysql_queries",
  "crates/lib/password",
+  "crates/lib/prompt_engineering",
  "crates/lib/redis_common",
  "crates/lib/redis_schema",
  "crates/lib/server_environment",
-  "crates/lib/shared_service_components",
  "crates/lib/sqlx_mysql_helpers",
  "crates/lib/storyteller_root",
  "crates/lib/subprocess_common",
@@ -121,40 +113,5 @@ members = [

  # cargo-hakari workspace hack (compilation speedup hack)
  # see usage: https://crates.io/crates/cargo-hakari
-  "crates/workspace-build-acceleration",
+  "crates/my-workspace-hack",
 ]
-
-[workspace.package]
-# Common inheritable settings for monorepo packages
-publish = false
-
-[workspace.dependencies]
-# Hakari workspace hack (faster compile times)
-workspace-build-acceleration = { path = "crates/workspace-build-acceleration" }
-
-# 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"
-chrono = { version = "0.4.40", features = ["serde"] }
-log = "0.4.27"
-once_cell = "1.21.3"
-rand = "0.9.0"
-regex = "1.11.1"
-reqwest = { version = "0.12.15", features = ["json", "cookies"] }
-serde = {  version = "1.0.219", features = ["derive"] }
-serde_derive = "1.0.219"
-serde_json = "1.0.140"
-serde_yaml = "0.9.1"
-toml = "0.7.4"
-url = "2.5.4"
-
-# Internal libs
-crockford = { path = "crates/lib/data/crockford" }
-errors = { path = "crates/lib/errors" }
-filesys = { path = "crates/lib/files/filesys" }
-memory_caching = { path = "crates/lib/caching/memory_caching" }
-shared_service_components = { path = "crates/lib/shared_service_components" }
-openai_sora_client = { path = "crates/lib/api_clients/openai_sora_client" }
-primitives = { path = "crates/lib/data/primitives" }
-subprocess_common = { path = "crates/lib/subprocess_common" }
@@ -13,7 +13,7 @@ name = "composite_identifiers"
 path = "src/lib.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -21,7 +21,7 @@ enums = { path = "../enums" }
 tokens = { path = "../tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }

 # Serialization
 serde = "1.0.125"
@@ -13,13 +13,13 @@ name = "enums"
 path = "src/lib.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }

 # Serialization
 serde = "1.0.125"
@@ -40,10 +40,6 @@ 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,
@@ -112,7 +108,6 @@ 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",
@@ -137,7 +132,6 @@ 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),
@@ -165,7 +159,6 @@ impl InferenceJobType {
      Self::GptSovits,
      Self::ComfyUi,
      Self::StudioGen2,
-      Self::ImageGenApi,
      Self::ConvertFbxToGltf,
      Self::MocapNet,
      Self::RvcV2,
@@ -204,7 +197,6 @@ 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");
@@ -228,7 +220,6 @@ 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");
@@ -252,7 +243,6 @@ 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);
@@ -271,7 +261,7 @@ mod tests {
    fn all_variants() {
      // Static check
      let mut variants = InferenceJobType::all_variants();
-      assert_eq!(variants.len(), 17);
+      assert_eq!(variants.len(), 16);
      assert_eq!(variants.pop_first(), Some(InferenceJobType::VideoRender));
      assert_eq!(variants.pop_first(), Some(InferenceJobType::LivePortrait));
      assert_eq!(variants.pop_first(), Some(InferenceJobType::FaceFusion));
@@ -279,7 +269,6 @@ 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));
@@ -33,10 +33,6 @@ 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")]
@@ -74,7 +70,6 @@ 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",
@@ -95,7 +90,6 @@ 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),
@@ -119,7 +113,6 @@ impl InferenceModelType {
      Self::VallEX,
      Self::RerenderAVideo,
      Self::StableDiffusion,
-      Self::ImageGenApi,
      Self::MocapNet,
      Self::StyleTTS2,
      Self::ConvertFbxToGltf,
@@ -147,7 +140,6 @@ 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");
@@ -167,7 +159,6 @@ 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");
@@ -185,7 +176,6 @@ 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);
@@ -197,18 +187,17 @@ mod tests {
    fn all_variants() {
      // Static check
      let mut variants = InferenceModelType::all_variants();
-      assert_eq!(variants.len(), 15);
+      assert_eq!(variants.len(), 13);
      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::SoVitsSvc));
      assert_eq!(variants.pop_first(), Some(InferenceModelType::SeedVc));
+      assert_eq!(variants.pop_first(), Some(InferenceModelType::SoVitsSvc));
      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));
@@ -57,9 +57,6 @@ pub enum MediaFileOriginModelType {
  #[serde(rename = "studio")]
  StorytellerStudio,

-  #[serde(rename = "studio_ig")]
-  StorytellerStudioImageGen,
-
  #[serde(rename = "vst")]
  VideoStyleTransfer,

@@ -96,7 +93,6 @@ impl MediaFileOriginModelType {
      Self::MocapNet => "mocap_net",
      Self::StyleTTS2 => "styletts2",
      Self::StableDiffusion15 => "stable_diffusion_1_5",
-      Self::StorytellerStudioImageGen => "studio_ig",
      Self::GptSovits => "gpt_sovits",
      Self::StorytellerStudio => "studio",
      Self::VideoStyleTransfer => "vst",
@@ -119,7 +115,6 @@ impl MediaFileOriginModelType {
      "mocap_net" => Ok(Self::MocapNet),
      "styletts2" => Ok(Self::StyleTTS2),
      "stable_diffusion_1_5" => Ok(Self::StableDiffusion15),
-      "studio_ig" => Ok(Self::StorytellerStudioImageGen),
      "gpt_sovits" => Ok(Self::GptSovits),
      "studio" => Ok(Self::StorytellerStudio),
      "vst" => Ok(Self::VideoStyleTransfer),
@@ -147,7 +142,6 @@ impl MediaFileOriginModelType {
      Self::StableDiffusion15,
      Self::GptSovits,
      Self::StorytellerStudio,
-      Self::StorytellerStudioImageGen,
      Self::VideoStyleTransfer,
      Self::ComfyUi,
      Self::VallEX,
@@ -13,11 +13,11 @@ name = "enums_public"
 path = "src/lib.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }

 # Internal
 enums = { path = "../enums" }
@@ -91,10 +91,6 @@ 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.
@@ -123,7 +119,6 @@ impl PublicInferenceJobType {
      InferenceJobType::Unknown => Self::Unknown,
      InferenceJobType::BevyToWorkflow => Self::BevyToWorkflow,
      InferenceJobType::RerenderAVideo => Self::RerenderAVideo,
-      InferenceJobType::ImageGenApi => Self::ImageGenApi,
    }
  }
  
@@ -145,7 +140,6 @@ 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,
@@ -64,10 +64,6 @@ pub enum PublicMediaFileModelType {
  #[serde(rename = "studio")]
  StorytellerStudio,

-  /// NB: This is GPT4o image generation
-  #[serde(rename = "studio_ig")]
-  StorytellerStudioImageGen,
-
  #[serde(rename = "vst")]
  VideoStyleTransfer,

@@ -93,7 +89,6 @@ impl PublicMediaFileModelType {
      MediaFileOriginModelType::LivePortrait => Self::FaceMirror,
      MediaFileOriginModelType::SadTalker => Self::FaceAnimator,
      MediaFileOriginModelType::StyleTTS2 => Self::VoiceDesigner,
-      MediaFileOriginModelType::StorytellerStudioImageGen => Self::StorytellerStudioImageGen,
      // Conserved variants
      MediaFileOriginModelType::RvcV2 => Self::RvcV2,
      MediaFileOriginModelType::SoVitsSvc => Self::SoVitsSvc,
@@ -118,7 +113,6 @@ impl PublicMediaFileModelType {
      Self::FaceAnimator => MediaFileOriginModelType::SadTalker,
      Self::Lipsync => MediaFileOriginModelType::FaceFusion,
      Self::VoiceDesigner => MediaFileOriginModelType::StyleTTS2,
-      Self::StorytellerStudioImageGen => MediaFileOriginModelType::StorytellerStudioImageGen,
      // Conserved variants
      Self::RvcV2 => MediaFileOriginModelType::RvcV2,
      Self::SoVitsSvc => MediaFileOriginModelType::SoVitsSvc,
@@ -13,7 +13,7 @@ name = "reusable_types"
 path = "src/lib.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]

@@ -21,7 +21,7 @@ normal = ["workspace-build-acceleration"]
 # (none)

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }

 # External
 chrono = { version = "0.4.22", features = ["serde"] }
@@ -13,7 +13,7 @@ name = "tokens"
 path = "src/lib.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]

@@ -21,7 +21,7 @@ normal = ["workspace-build-acceleration"]
 # (none)

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../my-workspace-hack" }

 # Serialization
 serde = "1.0.125"
@@ -13,7 +13,7 @@ name = "dev-database-seed"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -32,7 +32,7 @@ storyteller_root = { path = "../../../lib/storyteller_root" }
 tokens = { path = "../../../_types/tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 log = "0.4.14"
@@ -13,7 +13,7 @@ name = "dev-upload-media-file"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -29,7 +29,7 @@ mysql_queries = { path = "../../../lib/mysql_queries" }
 tokens = { path = "../../../_types/tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = "4.4.8"
@@ -13,7 +13,7 @@ name = "scrape-test"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]

@@ -22,7 +22,7 @@ errors = { path = "../../../lib/errors" }
 web_scrapers = { path = "../../../lib/web_scrapers" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = "4.4.8"
@@ -13,11 +13,11 @@ name = "db-backfill"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # Internal
 datetimes = { path = "../../../lib/data/datetimes" }
@@ -13,7 +13,7 @@ name = "db-backup"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -21,7 +21,7 @@ easyenv = { path = "../../../lib/easyenv" }
 errors = { path = "../../../lib/errors" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 anyhow = "1.0.86"
@@ -13,7 +13,7 @@ name = "db-cleanup"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -26,7 +26,7 @@ storyteller_root = { path = "../../../lib/storyteller_root" }
 tokens = { path = "../../../_types/tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = "4.4.8"
@@ -13,7 +13,7 @@ name = "elasticsearch-cli"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -29,7 +29,7 @@ storyteller_root = { path = "../../../lib/storyteller_root" }
 tokens = { path = "../../../_types/tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = "4.4.8"
@@ -13,7 +13,7 @@ name = "kubectl-pod-cleanup"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -22,7 +22,7 @@ easyenv = { path = "../../../lib/easyenv" }
 errors = { path = "../../../lib/errors" }

 # Workspace hack (faster compile yytimes)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 log = "0.4.20"
@@ -13,7 +13,7 @@ name = "lru-delete-models"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -21,7 +21,7 @@ errors = { path = "../../../lib/errors" }
 logging = { path = "../../../lib/logging" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = { version = "4.4.8", features = ["derive"] }
@@ -13,7 +13,7 @@ name = "migration-tool"
 path = "src/main.rs"

 [package.metadata.cargo-udeps.ignore]
-normal = ["workspace-build-acceleration"]
+normal = ["my-workspace-hack"]

 [dependencies]
 # Internal
@@ -30,7 +30,7 @@ storyteller_root = { path = "../../../lib/storyteller_root" }
 tokens = { path = "../../../_types/tokens" }

 # Workspace hack (faster compile times)
-workspace-build-acceleration = { version = "0.1", path = "../../../workspace-build-acceleration" }
+my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }

 # External
 clap = "4.4.8"
@@ -1,5 +1,5 @@
 [package]
-name = "tauri-realtime"
+name = "app"
 version = "0.1.0"
 description = "A Tauri App"
 authors = ["you"]
@@ -7,7 +7,6 @@ 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

@@ -16,64 +15,61 @@ name = "app_lib"
 crate-type = ["staticlib", "cdylib", "rlib"]

 [build-dependencies]
-tauri-build = { version = "2.1.1", features = [] }
+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 = "../../lib/machine_learning/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 = { version = "0.13.5", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
+cudarc = "0.13.7"

 # 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"
-chrono = "0.4.40"
-const_format = "0.2.34"
-directories = "6.0.0"
+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"
+directories = "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"
 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"
 serde_yaml = "0.9"
-tauri = { version = "2.4.1", features = [] }
-tauri-plugin-log = "2.3.1"
+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"] }

-# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
-
-[target.'cfg(not(target_os = "macos"))'.dependencies]
-ml_models = { path = "../../lib/machine_learning/ml_models" }
+[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"
-
-#[target.'cfg(target_os = "macos")'.features]
-#default  = []
-#ml_models = ["dep:ml_models"]

 [profile.dev]
 incremental = true # Compile your binary in smaller steps.
@@ -0,0 +1,16 @@
+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>,
+) -> Result<String, String> {
+  
+  download_all_models(&app_data_root).await
+    .map_err(|err| {
+      "error downloading models".to_string()
+    })?;
+  
+  Ok(("downloaded".to_string()))
+}
@@ -0,0 +1,146 @@
+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)
+}
@@ -0,0 +1,4 @@
+pub mod download_models;
+pub mod flip_image;
+pub mod image_endpoint;
+pub mod test_counter;
@@ -0,0 +1,24 @@
+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
+}
@@ -1,16 +1,13 @@
 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)]
@@ -29,17 +26,5 @@ 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,
-  },
+  }
 }
@@ -1,23 +1,17 @@
 pub mod endpoints;
 pub mod events;
+pub mod ml;
 pub mod state;
-pub mod stubs;
 pub mod threads;
 pub mod transfer;
-pub mod utils;

 use crate::endpoints::download_models::download_models;
-use crate::endpoints::flip_image::flip_image;
-use crate::endpoints::inpaint_image_endpoint::inpaint_image;
-use crate::endpoints::realtime_image_endpoint::infer_image;
-use crate::endpoints::remove_background_endpoint::remove_background;
-use crate::endpoints::save_image_endpoint::save_image;
-use crate::endpoints::text_to_image_endpoint::text_to_image;
+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::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;

@@ -45,15 +39,9 @@ pub fn run() {
  tauri::Builder::default()
    .plugin(tauri_plugin_log::Builder::new()
      .level(log::LevelFilter::Info)
-      .targets(vec![
-        Target::new(TargetKind::Stdout),
-        Target::new(TargetKind::LogDir { file_name: Some(app_data_root.log_file_name_str().to_string()) }),
-      ])
+      .targets(vec![Target::new(TargetKind::Stdout)])
      .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()
@@ -72,14 +60,9 @@ pub fn run() {
    .manage(model_cache)
    .manage(app_data_root)
    .invoke_handler(tauri::generate_handler![
-      download_models,
-      flip_image,
+      test_counter, 
      infer_image,
-      inpaint_image,
-      remove_background,
-      save_image,
-      text_to_image,
-      text_to_image,
+      download_models,
    ])
    .run(tauri::generate_context!())
    .expect("error while running tauri application");
@@ -0,0 +1,15 @@
+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(())
+}
@@ -0,0 +1,30 @@
+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)
+}
@@ -0,0 +1,2 @@
+pub mod download_model_registry_file;
+pub mod download_all_models;
@@ -0,0 +1,8 @@
+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;
@@ -1,14 +1,15 @@
+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<UNet2DConditionModel>>>,
+  unet2: RwLock<Option<Arc<UNetModel>>>,
 }

 impl ModelCache {
@@ -38,7 +39,7 @@ impl ModelCache {
    }
  }

-  pub fn set_unet(&self, vae: Arc<UNet2DConditionModel>) -> anyhow::Result<()> {
+  pub fn set_unet(&self, vae: Arc<UNetModel>) -> anyhow::Result<()> {
    match self.unet2.write() {
      Err(err) => Err(anyhow!("{}", err)),
      Ok(mut lock) => {
@@ -48,7 +49,7 @@ impl ModelCache {
    }
  }

-  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNet2DConditionModel>>> {
+  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNetModel>>> {
    match self.unet2.read() {
      Err(err) => Err(anyhow!("{}", err)),
      Ok(ref lock) => {
@@ -0,0 +1,86 @@
+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)
+  }
+}
@@ -0,0 +1,130 @@
+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(())
+  }
+ 
+}
@@ -0,0 +1,2 @@
+pub mod unet_model;
+pub mod lazy_load_vae_model;
@@ -0,0 +1,30 @@
+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))?)
+  }
+}
@@ -3,51 +3,19 @@ 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;
-
-
-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,
-}
+use crate::ml::model_type::ModelType;
+use crate::state::app_dir::AppWeightsDir;

 #[allow(clippy::too_many_arguments)]
-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;
-
+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> {
  info!("Preparing text embeddings... for {:?}", sd_version);

  let which = match sd_version {
@@ -71,8 +39,7 @@ pub fn infer_clip_text_embeddings<P: AsRef<Path>, Q: AsRef<Path>>(args: ClipArgs
        dtype,
        use_guide_scale,
        *first,
-        clip_json_path.as_ref(),
-        clip_weights_path.as_ref(),
+        weights_dir,
      )
    })
    .collect::<anyhow::Result<Vec<_>>>()?;
@@ -83,25 +50,20 @@ pub fn infer_clip_text_embeddings<P: AsRef<Path>, Q: AsRef<Path>>(args: ClipArgs
 }

 #[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, 
-  clip_json_path: &Path,
-  clip_weights_path: &Path,
-) -> 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, weights_dir: &AppWeightsDir) -> 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 = clip_json_path;
+  //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);

  info!("Loading Clip Tokenizer path: {:?}", tokenizer);

@@ -123,16 +85,14 @@ fn do_infer_clip_text_embeddings(
  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;

  println!("Building the Clip transformer.");
-  
-  // TODO(bt,2025-03-23): These used to be two different clip models. We should only need one.
+  let clip_weights_file = if first { ModelFile::Clip } else { ModelFile::Clip2 };
  let clip_weights = if first {
-    println!(">>>>> CLIP 1");
-    clip_weights_path
+    //clip_weights_file.get(clip_weights, sd_version, use_f16)?
+    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder)
  } else {
-    println!(">>>>> CLIP 2");
-    clip_weights_path
+    //clip_weights_file.get(clip2_weights, sd_version, use_f16)?
+    weights_dir.model_path(&ModelType::SdxlTurboClipEncoder2)
  };
-  
  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)?;

@@ -1,100 +1,72 @@
-use std::path::{Path, PathBuf};
+use std::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, Device, IndexOp, Tensor, D};
+use candle_core::{DType, IndexOp, Tensor, D};
 use candle_nn::VarBuilder;
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use candle_transformers::models::stable_diffusion::lcm::{LCMScheduler, LCMSchedulerConfig};
+use candle_transformers::models::stable_diffusion::lcm::LCMScheduler;
 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::{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 log::info;
 use rand::Rng;
+use tauri::{AppHandle, Emitter};

-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 struct Args<'a> {
  pub image: &'a DynamicImage,
  pub prompt: String,
  pub uncond_prompt: String,
  pub cfg_scale: Option<f64>,
-  pub img2img_strength: 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 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<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;
+pub fn lcm_pipeline(args: Args<'_>) -> Result<RgbImage> {
+  let Args { prompt, uncond_prompt, cfg_scale, i2i_strength, configs, model_cache, prompt_cache, app, image, app_data_root, use_flash_attn } = args;

-  let img2img_strength = img2img_strength
-    .map(|s| remap_lcm_strength_range(s))
-    .unwrap_or(DEFAULT_STRENGTH);
-  
-  let img2img_strength = img2img_strength / 100.0;
+  let weights_dir = app_data_root.weights_dir();

+  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(scheduler_steps, img2img_strength, LCMSchedulerConfig::default())?;
+  let mut scheduler = LCMScheduler::new(configs.scheduler_steps, img2img_strength, candle_transformers::models::stable_diffusion::lcm::LCMSchedulerConfig::default())?;

-  //let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
-  let seed = maybe_seed.unwrap_or(DEFAULT_SEED);
+  let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
+  configs.device.set_seed(seed)?;

-  device.set_seed(seed)?;
+  //info!("Using seed: {}", seed);

  let guidance_scale = match cfg_scale {
    Some(guidance_scale) => guidance_scale,
-    None => match sd_version {
+    None => match configs.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"
@@ -104,50 +76,72 @@ pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<Rg
  // 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!("Prompt is NOT cached! Calculating embedding...");
-    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
-      device,
-      dtype,
-      use_guide_scale,
-      clip_json_path,
-      clip_weights_path,
-    })?;
+    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,
+      use_f16, // use_f16
+      &configs.device,
+      configs.dtype,
+      use_guide_scale,
+      weights_dir,
+    )?;
    prompt_cache.store_copy(&prompt, &tensor)?;
    tensor
  };

-  debug!("Text embeddings shape: {:?}", text_embeddings.shape());
+  println!("Text embeddings shape: {:?}", text_embeddings.shape());

-  debug!("Loading input image into tensor...");
+  println!("Loading input image into tensor...");

-  let input_image = dynamic_image_to_tensor(image, &device, dtype)?;
+  let input_image = dynamic_image_to_tensor(image, &configs.device, configs.dtype)?;

-  debug!("Reference image shape: {:?}", input_image.shape());
+  println!("Reference image shape: {:?}", input_image.shape());

-  let vae_scale = get_vae_scale(sd_version);
+  let vae_scale = get_vae_scale(configs.sd_version);

  let maybe_vae = model_cache.get_vae()?;

  let vae = match maybe_vae {
    Some(vae) => vae,
    None => {
-      info!("Building VAE model from file {:?}...", vae_path.as_ref());
-      
-      let vae = sd_config.build_vae(vae_path, &device, dtype)?;
+      //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)?;

      let vae = Arc::new(vae);

@@ -162,9 +156,11 @@ pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<Rg
  let unet = match maybe_unet {
    Some(unet) => unet,
    None => {
-      info!("Building UNET model from file {:?}...", unet_path.as_ref());
+      info!("No unet found in cache; loading...");

-      let in_channels = match sd_version {
+      let unet_weights = weights_dir.model_path(&ModelType::SimianLuoLcmDreamshaperV7Unet);
+
+      let in_channels = match configs.sd_version {
        StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
        _ => 4,
      };
@@ -199,10 +195,12 @@ pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<Rg
      };

      // Load the model directly from the safetensors file
-      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_path], dtype, &device)? };
+      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_weights], configs.dtype, &configs.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())?;
@@ -213,40 +211,50 @@ pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<Rg

  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, &device, dtype)?;
+  let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &configs.device, configs.dtype)?;

  // Generate latents from input image using the LCM approach
-  debug!("Generating latents from input image...");
+  println!("Generating latents from input image...");
  let init_latents = (init_latent_dist.sample()? * vae_scale)?;
-
-  debug!("Initial latents shape: {:?}", init_latents.shape());
+  println!("Initial latents shape: {:?}", init_latents.shape());

  // Calculate timesteps for LCM with proper img2img handling
-  let timesteps = LCMScheduler::get_timesteps_for_steps(scheduler_steps, img2img_strength);
+  let timesteps = LCMScheduler::get_timesteps_for_steps(configs.scheduler_steps, img2img_strength);
  let t_start = timesteps[0]; // First denoising step

  // Add noise at the appropriate timestep for img2img
-  debug!("Adding noise to latents for t_start={}", t_start);
+  println!("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() {
-    info!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
+    println!("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, Some(&guidance_scale_embedding)) {
+    let mut noise_pred = match unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, &guidance_scale_embedding) {
      Ok(pred) => pred,
      Err(e) => {
-        error!("UNet inference failed with error: {}", e);
+        println!("UNet inference failed with error: {}", e);
        return Err(anyhow::anyhow!("UNet inference failed: {}", e));
      },
    };
@@ -258,22 +266,24 @@ pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<Rg
    }

    // Apply the LCM scheduler step
-    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, scheduler_steps)?;
+    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, configs.scheduler_steps)?;
  }

-  debug!("Diffusion process completed, decoding image...");
+  println!("Diffusion process completed, decoding image...");
  let image = vae.decode(&(latents / vae_scale)?)?;

-  debug!("Scaling image back");
+  println!("VAE decode completed");
+
+  println!("Scaling image back");
  let image = ((image / 2.)? + 0.5)?;

-  debug!("Normalized image values");
+  println!("Normalized image values");
  let image = (image.clamp(0f32, 1.)? * 255.)?;

-  debug!("Convert to int8");
+  println!("Convert to int8");
  let image = image.to_dtype(DType::U8)?;

-  debug!("Converted to 8-bit format");
+  println!("Converted to 8-bit format");

  let image = tensor_to_image_buffer(&image.i(0)?)?;

@@ -1,5 +1,4 @@
 pub mod get_vae_scale;
 pub mod infer_clip_text_embeddings;
 pub mod stable_diffusion_pipeline;
-pub mod lcm_pipeline;
-mod remap_lcm_strength_range;
+pub mod lcm_pipeline;
@@ -1,78 +1,76 @@
-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, ClipArgs};
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
+use crate::state::app_config::AppConfig;
 use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, Device, IndexOp, Tensor, D};
+use candle_core::{DType, 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, P: AsRef<Path>, Q: AsRef<Path>, R: AsRef<Path>, S: AsRef<Path>> {
+pub struct Args<'a> {
    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 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 app_data_root: &'a AppDataRoot,
+    pub app: &'a AppHandle,
 }

-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>
-{
+pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
    let Args { 
        prompt, 
        uncond_prompt, 
        cfg_scale, 
        i2i_strength, 
+        configs, 
        model_cache, 
        prompt_cache, 
+        app, 
        image,
-        sd_version,
-        sd_config,
-        device,
-        dtype,
-        maybe_seed,
-        scheduler_steps,
-        sdxl_turbo_vae_path,
-        sdxl_turbo_unet_path,
-        clip_json_path,
-        clip_weights_path,
+        app_data_root,
    } = 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 = sd_config.build_scheduler(scheduler_steps)?;
+    let mut scheduler = configs.sd_config.build_scheduler(
+        configs.scheduler_steps)?;

-    let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
-    
-    device.set_seed(seed)?;
+    let seed = configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
+    configs.device.set_seed(seed)?;

    info!("Using seed: {}", seed);

    let guidance_scale = match cfg_scale {
        Some(guidance_scale) => guidance_scale,
-        None => match sd_version {
+        None => match configs.sd_version {
            StableDiffusionVersion::V1_5
            | StableDiffusionVersion::V2_1
            | StableDiffusionVersion::Xl => 7.5,
@@ -93,21 +91,20 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
    } else {
        
        info!("Prompt is NOT cached! Calculating embedding...");
-        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,
+        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,
            use_guide_scale,
-            clip_json_path: clip_json_path.as_ref(),
-            clip_weights_path: clip_weights_path.as_ref(),
-        })?;
+            weights_dir,
+        )?;
        prompt_cache.store_copy(&prompt, &tensor)?;
        tensor
    };
@@ -116,11 +113,14 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)

    println!("Loading input image into tensor...");

-    let input_image = dynamic_image_to_tensor(image, device, dtype)?;
+    let input_image = dynamic_image_to_tensor(
+        image,
+        &configs.device,
+        configs.dtype)?;

    println!("Reference image shape: {:?}", input_image.shape());

-    let vae_scale = get_vae_scale(sd_version);
+    let vae_scale = get_vae_scale(configs.sd_version);

    let maybe_vae = model_cache.get_vae()?;
    
@@ -129,14 +129,33 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
        None => {
            info!("No vae found in cache; loading...");
            
-            let vae_file = sdxl_turbo_vae_path.as_ref();
+            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 = sd_config.build_vae(vae_file, &device, dtype)?;
+            let vae_file = weights_dir.model_path(&ModelType::SdxlTurboVae);
+            
+            let vae = configs
+              .sd_config
+              .build_vae(vae_file, &configs.device, configs.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
        }
    };
@@ -148,15 +167,32 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
        None => {
            info!("No unet found in cache; loading...");

-            let unet_weights_file = sdxl_turbo_unet_path.as_ref();
+            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 = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)
+            let unet = UNetModel::new(&configs.sd_config, unet_file, &configs.device, configs.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
        }
    };
@@ -180,9 +216,9 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
    };

    let t_start = {
-        let start = scheduler_steps - (scheduler_steps as f64 * img2img_strength) as usize;
+        let start = configs.scheduler_steps - (configs.scheduler_steps as f64 * img2img_strength) as usize;

-        println!("Starting from step {} of {} (strength: {})", start, scheduler_steps, img2img_strength);
+        println!("Starting from step {} of {} (strength: {})", start, configs.scheduler_steps, img2img_strength);
        start
    };

@@ -221,7 +257,7 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)

        let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep)?;

-        let mut noise_pred = match unet.forward(&latent_model_input, timestep as f64, &text_embeddings) {
+        let mut noise_pred = match unet.inference(&latent_model_input, timestep as f64, &text_embeddings) {
            Ok(pred) => pred,
            Err(e) => {
                println!("UNet inference failed with error: {}", e);
@@ -1,38 +1,62 @@
 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;

 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
+/// 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,
+  pub scheduler_samples: usize,
+  
+  pub seed: Option<u64>,
+
+  pub cfg_scale: Option<f64>,
+  
  /// Probably shouldn't be here.
  pub hf_api: Api,
+  
+  /// Root location for application data.
+  pub app_data_root: AppDataRoot,
 }

-impl ModelConfig {
+impl AppConfig {
+  
  pub fn init() -> anyhow::Result<Self> {
+    let yaml_configs = YamlConfig::load_from_config_file_recursive();
+    
+    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))
@@ -47,15 +71,39 @@ impl ModelConfig {
        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 {
+      OsPlatform::Linux => yaml_configs.linux_default_data_path.as_deref(),
+      OsPlatform::MacOs => yaml_configs.mac_default_data_path.as_deref(),
+      OsPlatform::Windows => yaml_configs.windows_default_data_path.as_deref(),
+    };
+    
+    println!("Possible user-defined data root: {:?}", default_data_root);
+
+    let app_data_root = match default_data_root {
+      None => AppDataRoot::create_default()?,
+      Some(path) => AppDataRoot::create_existing(path)?,
+    };
+
+    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,
    })
  }
-}
+}
@@ -1,21 +1,17 @@
+use crate::ml::model_type::ModelType;
 use crate::state::expanduser::expanduser;
 use crate::state::os_platform::OsPlatform;
 use anyhow::anyhow;
 use directories::UserDirs;
 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 {
@@ -23,8 +19,6 @@ pub struct AppDataRoot {
  assets_dir: AppAssetsDir,
  weights_dir: AppWeightsDir,
  temp_dir: TemporaryDir,
-  log_file_name: PathBuf,
-  log_file_name_string: String,
 }

 #[derive(Clone)]
@@ -75,19 +69,12 @@ 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,
    })
  }
  
@@ -106,18 +93,6 @@ 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
-  }
-
-  pub fn get_sora_cookie_file_path(&self) -> PathBuf {
-    self.path.join("sora_cookies.txt")
-  }
 }

 impl AppAssetsDir {
@@ -141,18 +116,6 @@ 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 {
@@ -176,6 +139,10 @@ 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())
+  }
 }

 impl TemporaryDir {
@@ -1,11 +1,10 @@
 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::StreamExt;
 use futures::stream::FuturesUnordered;
+use futures::StreamExt;
 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;
@@ -13,7 +12,6 @@ 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;
@@ -41,32 +39,22 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !

  let mut download_queue = VecDeque::new();

-  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);
+  // 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);

  let mut handles = FuturesUnordered::new();

  while let Some(model) = download_queue.pop_front() {
-    let url = model.r2_download_url.to_string();
+    let url = model.get_download_url().to_string();

-    let filename = app_data_root.weights_dir().weight_path(&model);
+    let filename = app_data_root.weights_dir().model_path(&model);

    if filename.exists() {
      info!("File already exists: {:?}", &filename);
@@ -87,18 +75,9 @@ 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.name,
-      model_type,
+      model_name: model.get_name(),
+      model_type: model.get_notification_type(),
    });

    if let Err(err) = result {
@@ -123,8 +102,8 @@ pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> !
        lock.push(ModelDownloadStatusChannel {
          receiver: rx,
          final_file_path: filename.clone(),
-          model_name: model.to_descriptive_name(),
-          model_type,
+          model_name: model.get_name().to_string(),
+          model_type: model.get_notification_type(),
        })
      }
    }
@@ -1,74 +0,0 @@
-[package]
-name = "artcraft"
-version = "0.1.0"
-description = "Artcraft"
-authors = ["you"]
-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
-
-[lib]
-name = "app_lib"
-crate-type = ["staticlib", "cdylib", "rlib"]
-
-[build-dependencies]
-tauri-build = { version = "2.1.1", features = [] }
-
-[dependencies]
-# Workspace hack (faster compile times)
-workspace-build-acceleration = { workspace =  true }
-
-# Internal
-crockford = { workspace = true }
-errors = { workspace = true }
-filesys = { workspace = true }
-memory_caching = { workspace = true }
-openai_sora_client = { workspace = true }
-primitives = { workspace = true }
-subprocess_common = { workspace = true }
-
-# External
-#openssl = { version = "0.10", features = ["vendored"] }
-# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
-anyhow = { workspace = true }
-base64 = "0.22.1"
-bytes = "1.10"
-chrono = { workspace = true }
-const_format = "0.2.34"
-directories = "6.0.0"
-dirs = "6.0.0"
-futures = "0.3"
-image = "0.25.5"
-log = { workspace = true }
-once_cell = { workspace = true }
-rand = { workspace = true }
-reqwest = { workspace = true, features = ["multipart", "stream"], default-features = false }
-serde = { workspace = true }
-serde_derive = { workspace = true }
-serde_json = { workspace = true }
-serde_yaml = { workspace = true }
-tauri = { version = "2.4.1", features = ["unstable"] }
-tauri-plugin-log = "2.3.1"
-tempdir = "0.3.7"
-tempfile = "3.18.0"
-tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
-tokio-util = { version = "0.7", features = ["codec"] }
-
-[dev-dependencies]
-libloading = "0.8.6"
-serial_test = "3.2.0"
-assertor = "0.0.3"
-
-[profile.dev]
-incremental = true # Compile your binary in smaller steps.
-
-[profile.release]
-codegen-units = 1 # Allows LLVM to perform better optimization.
-lto = true # Enables link-time-optimizations.
-opt-level = 3 # `s` prioritizes small binary size. Use `3` if you prefer speed.
-panic = "abort" # Higher performance by disabling panic handlers.
-strip = true # Ensures debug symbols are removed.
@@ -1,90 +0,0 @@
-use std::fs::read_to_string;
-use base64::prelude::BASE64_STANDARD;
-use errors::AnyhowResult;
-use base64::Engine;
-use image::codecs::png::{CompressionType, FilterType, PngEncoder};
-use image::{DynamicImage, ImageReader};
-use std::io::Cursor;
-use log::{error, info};
-use tauri::{AppHandle, Manager, State};
-use openai_sora_client::credentials::SoraCredentials;
-use openai_sora_client::image_gen::common::{ImageSize, NumImages};
-use openai_sora_client::image_gen::sora_image_gen_remix::{sora_image_gen_remix, SoraImageGenRemixRequest};
-use openai_sora_client::upload::upload_media_from_bytes::sora_media_upload_from_bytes;
-use openai_sora_client::upload::upload_media_from_file::SoraMediaUploadRequest;
-use crate::state::app_dir::AppDataRoot;
-
-#[tauri::command]
-pub async fn image_generation_command(
-  image: &str,
-  prompt: &str,
-  app: AppHandle,
-  app_data_root: State<'_, AppDataRoot>,
-) -> Result<String, String> {
-  info!("image_generation_command called; processing image...");
-
-  let bytes = BASE64_STANDARD.decode(image)
-    .map_err(|err| format!("Base64 decode error: {}", err))?;
-
-  generate_image(bytes, prompt, &app_data_root)
-    .await
-    .map_err(|err| {
-      error!("error: {:?}", err);
-      "there was an error".to_string()
-    })?;
-
-  Ok("success".to_string())
-}
-
-pub async fn generate_image(file_bytes: Vec<u8>, prompt: &str, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
-  let sora_credentials = get_credentials(app_data_root)?;
-
-  let filename = "image.png".to_string();
-
-  let response= sora_media_upload_from_bytes(file_bytes, filename, &sora_credentials)
-      .await
-      .map_err(|err| {
-        error!("Failed to upload image to Sora: {:?}", err);
-        err
-      })?;
-
-  let sora_media_tokens = vec![response.id];
-
-  let response = sora_image_gen_remix(SoraImageGenRemixRequest {
-    prompt: prompt.to_string(),
-    num_images: NumImages::One,
-    image_size: ImageSize::Square,
-    sora_media_tokens: sora_media_tokens.clone(),
-    credentials: &sora_credentials,
-  }).await
-      .map_err(|err| {
-        error!("Failed to call Sora image generation: {:?}", err);
-        err
-      })?;
-
-  println!(">> TASK ID: {:?} ", response.task_id);
-
-  Ok(())
-}
-
-fn get_credentials(app_data_root: &AppDataRoot) -> AnyhowResult<SoraCredentials> {
-  let cookie_file = app_data_root.get_sora_cookie_file_path();
-
-  let bearer = read_to_string("/Users/bt/dev/storyteller/storyteller-rust/test_data/temp/bearer.txt")?
-      .trim()
-      .to_string();
-
-  let cookie= read_to_string(cookie_file)?
-      .trim()
-      .to_string();
-
-  let sentinel = read_to_string("/Users/bt/dev/storyteller/storyteller-rust/test_data/temp/sentinel.txt")?
-      .trim()
-      .to_string();
-
-  Ok(SoraCredentials {
-    bearer_token: bearer,
-    cookie: cookie,
-    sentinel: Some(sentinel),
-  })
-}
@@ -1,3 +0,0 @@
-pub mod flip_image;
-pub mod image_generation_command;
-pub mod open_login_command;
@@ -1,72 +0,0 @@
-use std::time::Duration;
-use log::{error, info};
-use once_cell::sync::Lazy;
-use tauri::{AppHandle, Manager, Url, WebviewUrl, WebviewWindowBuilder, WindowBuilder};
-use tokio::time::sleep_until;
-use errors::AnyhowResult;
-use crate::threads::login_thread::{LOGIN_WINDOW_NAME, SORA_LOGIN_URL};
-use crate::utils::clear_all_webview_cookies::clear_all_webview_cookies;
-
-pub const START_URL_STR: &str = "https://storyteller.ai/";
-
-pub static START_URL: Lazy<Url> = Lazy::new(|| {
-  Url::parse(START_URL_STR).expect("URL should parse")
-});
-
-pub const SORA_HOMEPAGE_URL_STR: &str = "https://sora.com/";
-
-pub static SORA_HOMEPAGE_URL : Lazy<Url> = Lazy::new(|| {
-  Url::parse(SORA_HOMEPAGE_URL_STR).expect("URL should parse")
-});
-
-#[tauri::command]
-pub async fn open_login_command(
-  app: AppHandle,
-) -> Result<String, String> {
-  info!("open_login_command called");
-
-  do_open_login(&app)
-    .await
-    .map_err(|err| {
-      error!("Error opening login: {:?}", err);
-      format!("Error opening login: {:?}", err)
-    })?;
-
-  Ok("result".to_string())
-}
-
-async fn do_open_login(app: &AppHandle) -> AnyhowResult<()> {
-  info!("Building login window...");
-
-  let url = WebviewUrl::External(START_URL.clone());
-
-  let window = WebviewWindowBuilder::new(app, LOGIN_WINDOW_NAME, url)
-      //.user_agent(openai_sora_client::credentials::USER_AGENT)
-      .always_on_top(true)
-      .title("Login to OpenAI")
-      .center()
-      .resizable(true)
-      .visible(true)
-      .closable(true)
-      .focused(true)
-      .devtools(true)
-      .build()?;
-
-  for (name, webview) in window.webviews() {
-    if name != LOGIN_WINDOW_NAME {
-      continue;
-    }
-
-    clear_all_webview_cookies(&webview)?;
-
-    // NB: We're starting to get Cloudflare protection screens. Let's try to avoid.
-    webview.navigate(SORA_HOMEPAGE_URL.clone())?;
-    tokio::time::sleep(Duration::from_millis(100)).await;
-
-    webview.navigate(SORA_LOGIN_URL.clone())?;
-    break;
-  }
-
-  info!("Done.");
-  Ok(())
-}
@@ -1,65 +0,0 @@
-pub mod commands;
-pub mod events;
-pub mod state;
-pub mod threads;
-pub mod transfer;
-pub mod utils;
-
-use crate::commands::flip_image::flip_image;
-use crate::commands::image_generation_command::image_generation_command;
-use crate::commands::open_login_command::open_login_command;
-use crate::state::app_config::AppConfig;
-use crate::threads::login_thread::login_thread;
-
-use tauri_plugin_log::Target;
-use tauri_plugin_log::TargetKind;
-
-#[cfg_attr(mobile, tauri::mobile_entry_point)]
-pub fn run() {
-  // NB: Tauri wants to install the logger itself, so we can't rely on the logger crate 
-  // until the tauri runtime begins.
-  println!("Loading model config...");
-
-  let config = AppConfig::init()
-    .expect("config should load");
-
-  let app_data_root = config.app_data_root.clone();
-  let app_data_root2 = config.app_data_root.clone();
-
-  println!("Initializing backend runtime...");
-
-  tauri::Builder::default()
-    .plugin(tauri_plugin_log::Builder::new()
-      .level(log::LevelFilter::Info)
-      .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()
-      //      .level(log::LevelFilter::Info)
-      //      .build(),
-      //  )?;
-      //}
-      let app = app.handle().clone();
-
-      tauri::async_runtime::spawn(login_thread(app, app_data_root2));
-
-      Ok(())
-    })
-    .manage(config)
-    .manage(app_data_root)
-    .invoke_handler(tauri::generate_handler![
-      flip_image,
-      image_generation_command,
-      open_login_command,
-    ])
-    .run(tauri::generate_context!("tauri.conf.json"))
-    .expect("error while running tauri application");
-}
@@ -1,6 +0,0 @@
-// Prevents additional console window on Windows in release, DO NOT REMOVE!!
-#![cfg_attr(not(debug_assertions), windows_subsystem = "windows")]
-
-fn main() {
-  app_lib::run();
-}
@@ -1,58 +0,0 @@
-use crate::state::app_dir::AppDataRoot;
-use crate::state::os_platform::OsPlatform;
-use crate::state::yaml_config::YamlConfig;
-
-const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
-const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
-
-
-/// A central place to configure app-wide details.
-pub struct AppConfig {
-  pub image_height: usize,
-  pub image_width: usize,
-  pub scheduler_steps: usize,
-  pub scheduler_samples: usize,
-  
-  pub seed: Option<u64>,
-
-  pub cfg_scale: Option<f64>,
-  
-  /// Root location for application data.
-  pub app_data_root: AppDataRoot,
-}
-
-impl AppConfig {
-  
-  pub fn init() -> anyhow::Result<Self> {
-    let yaml_configs = YamlConfig::load_from_config_file_recursive();
-    
-    println!("Configs: {:?}", yaml_configs);
-    
-    let os_platform = OsPlatform::get();
-
-    let default_data_root = match os_platform {
-      OsPlatform::Linux => yaml_configs.linux_default_data_path.as_deref(),
-      OsPlatform::MacOs => yaml_configs.mac_default_data_path.as_deref(),
-      OsPlatform::Windows => yaml_configs.windows_default_data_path.as_deref(),
-    };
-    
-    println!("Possible user-defined data root: {:?}", default_data_root);
-
-    let app_data_root = match default_data_root {
-      None => AppDataRoot::create_default()?,
-      Some(path) => AppDataRoot::create_existing(path)?,
-    };
-
-    println!("App data root: {:?}", app_data_root.path());
-
-    Ok(Self {
-      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,
-      app_data_root,
-    })
-  }
-}
@@ -1,103 +0,0 @@
-use std::fs::OpenOptions;
-use std::io::Write;
-use anyhow::anyhow;
-use log::{error, info};
-use once_cell::sync::Lazy;
-use reqwest::Url;
-use tauri::{AppHandle, Manager, Webview};
-use errors::AnyhowResult;
-use crate::state::app_dir::AppDataRoot;
-use crate::utils::sora_webview_cookies::get_all_sora_cookies_as_string;
-
-pub const LOGIN_WINDOW_NAME: &str = "login_window";
-
-pub const SORA_LOGIN_URL_STR: &str = "https://sora.com/auth/login?callback_path=/";
-
-pub static SORA_LOGIN_URL: Lazy<Url> = Lazy::new(|| {
-  Url::parse(SORA_LOGIN_URL_STR).expect("URL should parse")
-});
-
-pub const SORA_ROOT_URL_STR: &str = "https://sora.com/";
-
-pub static SORA_ROOT_URL: Lazy<Url> = Lazy::new(|| {
-  Url::parse(SORA_ROOT_URL_STR).expect("URL should parse")
-});
-
-pub async fn login_thread(app: AppHandle, app_data_root: AppDataRoot) -> ! {
-  loop {
-    for (window_name, webview) in app.webviews() {
-      if window_name == LOGIN_WINDOW_NAME {
-        let result = check_login_window(&webview, &app_data_root).await;
-        if let Err(err) = result {
-          error!("Error checking login window: {:?}", err);
-        }
-        break;
-      }
-    }
-    tokio::time::sleep(std::time::Duration::from_millis(2_000)).await;
-  }
-}
-
-async fn check_login_window(webview: &Webview, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
-  clear_browsing_data_on_test_domain(webview)?;
-  //keep_on_task(webview)?;
-  extract_cookies_to_file(webview, app_data_root)?;
-  // TODO: Other maintenance tasks ...
-  Ok(())
-}
-
-fn keep_on_task(webview: &Webview) -> AnyhowResult<()> {
-  let url = webview.url()?;
-  let hostname= url.host()
-      .ok_or(anyhow!("no host in url"))?
-      .to_string();
-  match hostname.as_str() {
-    "auth.openai.com" => {},
-    "openai.com" => {},
-    "sora.com" => {},
-    // Third party SSO
-    "accounts.google.com" => {},
-    "accounts.youtube.com" => {},
-    "login.live.com" => {},
-    "appleid.apple.com" => {},
-    _ => {
-      info!("Non login hostname: {}", hostname);
-      webview.navigate(SORA_LOGIN_URL.clone())?;
-    }
-  }
-  Ok(())
-}
-
-/// This is just so we have a way to clear browsing data.
-fn clear_browsing_data_on_test_domain(webview: &Webview) -> AnyhowResult<()> {
-  let url = webview.url()?;
-  let hostname= url.host()
-      .ok_or(anyhow!("no host in url"))?
-      .to_string();
-  match hostname.as_str() {
-    "storyteller.ai" => {
-      info!("Clearing all browsing data...");
-      webview.clear_all_browsing_data()?;
-    }
-    _ => {}
-  }
-  Ok(())
-}
-
-// TODO(bt,2025-04-07): Heuristic to detect when logged in. Only write when logged in.
-fn extract_cookies_to_file(webview: &Webview, app_data_root: &AppDataRoot) -> AnyhowResult<()> {
-  let cookies = get_all_sora_cookies_as_string(webview)?;
-
-  let cookie_file = app_data_root.get_sora_cookie_file_path();
-
-  let mut file = OpenOptions::new()
-      .create(true)
-      .write(true)
-      .truncate(true)
-      .open(cookie_file)?;
-
-  file.write_all(cookies.as_bytes())?;
-  file.flush()?;
-
-  Ok(())
-}
@@ -1 +0,0 @@
-pub mod login_thread;
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Brandon Thomas	b2a7a40f5f	don't download sdxl anymore	2025-03-13 00:31:51 -04:00
Brandon Thomas	9085c288bd	stop downloading clip tokenizer from hf	2025-03-13 00:24:06 -04:00
Brandon Thomas	dd71d022ae	vae wasn't cached in memory	2025-03-12 23:58:52 -04:00
Brandon Thomas	692c409e1b	load text encoder from our bucket	2025-03-12 23:54:42 -04:00
Brandon Thomas	3ed775cf76	vae and unet from cloudflare instead of hf	2025-03-12 23:38:49 -04:00
Brandon Thomas	08ea29d44c	cleanup lcm	2025-03-12 22:16:18 -04:00