Feature/smooth audio waveforms

.gitignore

@@ -84,6 +84,10 @@ backend/generated/
 # Build artifacts
 release/
 python-embed/
+# Generated at build time (sha256 of `uv export`); bundled into the app so the
+# first-run downloader can match the hosted python-embed archive. Regenerated
+# each build — must not be committed, or it would go stale vs. the lockfile.
+python-deps-hash.txt
 
 # Downloaded during CI for Windows NSIS installer
 resources/vc_redist.x64.exe

NOTICES.md

@@ -37,12 +37,38 @@ used by LTX Desktop.
   License: Apache License 2.0
   https://huggingface.co/hr16/yolox-onnx
 
+- **Krea 2 Turbo**
+  Copyright (c) Krea AI.
+  License: Krea 2 Community License
+  https://huggingface.co/krea/Krea-2-Turbo
+
+- **Qwen-Image-Edit-2511**
+  Copyright (c) Alibaba Cloud / Qwen Team.
+  License: Apache License 2.0
+  https://huggingface.co/Qwen/Qwen-Image-Edit-2511
+
+- **Qwen-Image-Edit-2511 (GGUF quantization)**
+  Copyright (c) Unsloth AI; base model (c) Alibaba Cloud / Qwen Team.
+  License: Apache License 2.0
+  https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF
+
+- **Qwen-Image-Edit-2511 Multiple Angles LoRA**
+  Copyright (c) fal.ai.
+  License: Apache License 2.0
+  https://huggingface.co/fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA
+
+- **Qwen-Image-Edit-2511 Lightning LoRA**
+  Copyright (c) LightX2V contributors.
+  License: Apache License 2.0
+  https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning
+
 ---
 
 ## Python Dependencies
 
 - **pillow** — Copyright (c) Jeffrey A. Clark and Pillow contributors — HPND License
 - **torch** — Copyright (c) Meta Platforms, Inc. — BSD 3-Clause License
+- **torchvision** — Copyright (c) Soumith Chintala and torchvision contributors — BSD 3-Clause License
 - **huggingface-hub** — Copyright (c) Hugging Face — Apache License 2.0
 - **tqdm** — Copyright (c) Casper da Costa-Luis — Mozilla Public License 2.0
 - **pynvml** — Copyright (c) NVIDIA Corporation — BSD 3-Clause License
@@ -57,6 +83,8 @@ used by LTX Desktop.
 - **protobuf** — Copyright (c) Google LLC — BSD 3-Clause License
 - **transformers** — Copyright (c) Hugging Face — Apache License 2.0
 - **sentencepiece** — Copyright (c) Google LLC — Apache License 2.0
+- **bitsandbytes** — Copyright (c) Facebook, Inc. and its affiliates — MIT License
+- **gguf** — Copyright (c) GGML (ggml.ai) — MIT License
 - **sageattention** — Copyright (c) Jintao Zhang et al. — Apache License 2.0
 - **natten** — Copyright (c) Ali Hassani, Steven Walton, et al. (SHI Labs) — Apache License 2.0
 - **opencv-python-headless** — Copyright (c) OpenCV team — Apache License 2.0
@@ -77,8 +105,11 @@ used by LTX Desktop.
 - **electron-updater** — Copyright (c) electron-userland — MIT License
 - **js-yaml** — Copyright (c) Vitaly Puzrin — MIT License
 - **lucide-react** — Copyright (c) Lucide contributors — ISC License
-- **react-dropzone** — Copyright (c) Param Aggarwal — MIT License
+- **react-resizable-panels** — Copyright (c) Brian Vaughn — MIT License
 - **tailwind-merge** — Copyright (c) dcastil — MIT License
+- **use-sync-external-store** — Copyright (c) Meta Platforms, Inc. and affiliates — MIT License
+- **zod** — Copyright (c) Colin McDonnell — MIT License
+- **zustand** — Copyright (c) Paul Henschel and Poimandres contributors — MIT License
 
 ---
 

README.md

@@ -148,6 +148,45 @@ Your models folder is the location you chose during setup, or the default for yo
 
 Open the selector to tick one or more LoRAs and adjust each one's strength. If you added files while the app was already open, reopen the generation view (or restart the app) — the installed‑LoRA list is read on load.
 
+## LoRA/IC-LoRA Library
+
+LoRA adapters let you steer **local** video generation toward a specific style or subject. They apply to local generation (Windows/Linux NVIDIA hardware or Apple Silicon Macs), for text-to-video, image-to-video, and audio-to-video — not in API/cloud mode.
+
+In local mode, **Browse LoRAs** (styles/subjects for text/image/audio-to-video) and **Browse IC-LoRAs** (in-context effects, e.g. video-to-video) open a built-in library of ready-to-download LoRAs — no manual file placement needed. Each entry shows a preview, its instructions, and a link to its Hugging Face page; some gated models require signing in with Hugging Face first. Entries authored by **LTX** are official; everything else is community-contributed and flagged with a disclaimer that LTX doesn't endorse or take responsibility for it.
+
+### Custom LoRAs
+
+You can also use your own `.safetensors` LoRA files instead of (or alongside) the library.
+
+**Which LoRAs are supported.** The app generates locally with the **LTX‑2.3 (22B)** model. "LTX‑2" is the model generation and "2.3" is its current release, so LoRAs labeled **LTX‑2** or **LTX‑2.3** both target this model and are supported. This includes LoRAs exported from ComfyUI for LTX‑Video (their key names are remapped automatically). A LoRA whose tensors don't match the model is simply skipped, so it has no effect rather than producing an error. LoRAs built for other base models (e.g. SDXL, Wan, Hunyuan, or older LTXV 0.9.x) target a different architecture and will not take effect.
+
+**Where to put the files.** Library downloads land in their own subfolder under `loras/` (one per LoRA), so it's safe to drop your own `.safetensors` files straight into the `loras/` (or `lora/`) subfolder of your models folder:
+
+```
+models/
+└── loras/
+    ├── cinematic.safetensors
+    ├── claymation.safetensors
+    └── <lora-id>/              ← library download, one subfolder per LoRA
+        └── model.safetensors
+```
+
+> **Use the `loras/` subfolder.** A file is detected as a LoRA only if it lives in a folder named `loras`/`lora` **or** its filename contains `lora` — and many LoRAs aren't named that way. Dropping them straight into `loras/` is the reliable option; it won't collide with library downloads, which each get their own subfolder. Subfolders are scanned recursively, so further nesting is fine.
+
+Your models folder is the location you chose during setup, or the default for your platform:
+
+- **Windows:** `%LOCALAPPDATA%\LTXDesktop\models\`
+- **macOS:** `~/Library/Application Support/LTXDesktop/models/`
+- **Linux:** `$XDG_DATA_HOME/LTXDesktop/models/` (default: `~/.local/share/LTXDesktop/models/`)
+
+**When the LoRA button appears.** In the generation view, a **LoRA** selector appears next to the other video settings (FPS, aspect ratio) only when **all** of these are true:
+
+1. You're in **Video** generation mode (not Image, Retake, or IC‑LoRA).
+2. You're generating **locally** — not in API/cloud generation mode.
+3. **At least one** LoRA file is detected in the models folder.
+
+Open the selector to tick one or more LoRAs and adjust each one's strength. If you added files while the app was already open, reopen the generation view (or restart the app) — the installed‑LoRA list is read on load.
+
 ## API keys, cost, and privacy
 
 ### LTX API key

backend/_routes/qwen_multiangle.py

@@ -0,0 +1,18 @@
+"""Route handler for POST /api/qwen-multiangle/generate."""
+
+from __future__ import annotations
+
+from fastapi import APIRouter, Depends
+
+from api_types import QwenMultiAngleGenerateRequest, QwenMultiAngleGenerateResponse
+from state import get_state_service
+from app_handler import AppHandler
+
+router = APIRouter(prefix="/api", tags=["qwen-multiangle"])
+
+
+@router.post("/qwen-multiangle/generate", response_model=QwenMultiAngleGenerateResponse)
+def route_qwen_multiangle_generate(
+    req: QwenMultiAngleGenerateRequest, handler: AppHandler = Depends(get_state_service)
+) -> QwenMultiAngleGenerateResponse:
+    return handler.qwen_multiangle.generate(req)

backend/api_types.py

@@ -30,7 +30,9 @@
     "gemma4-12b-with-proj-ltx-2.5",
     "gemma-4-e2b-it",
     "z-image-turbo",
+    "krea-2-turbo",
 ]
+ImageGenerationModelCheckpointID = Literal["z-image-turbo", "krea-2-turbo"]
 LTXLocalModelId = Literal[
     "ltx-2.5-22b-distilled",
     "ltx-2.3-22b-distilled-1.1",
@@ -161,6 +163,8 @@ class SuggestGapPromptResponse(BaseModel):
 class GenerateVideoCompleteResponse(BaseModel):
     status: Literal["complete"]
     video_path: str
+    # Seed actually used (None for API providers that don't expose it).
+    seed: int | None = None
 
 
 class GenerateVideoCancelledResponse(BaseModel):
@@ -173,6 +177,8 @@ class GenerateVideoCancelledResponse(BaseModel):
 class GenerateImageCompleteResponse(BaseModel):
     status: Literal["complete"]
     image_paths: list[str]
+    # Base seed; image i in the batch used seed + i.
+    seed: int | None = None
 
 
 class GenerateImageCancelledResponse(BaseModel):
@@ -475,7 +481,7 @@ class GenerateVideoRequest(BaseModel):
     lastImagePath: str | None = None
     keyframes: list[KeyframeInput] = Field(default_factory=list[KeyframeInput])
     audioPath: str | None = None
-    aspectRatio: Literal["16:9", "9:16"] = "16:9"
+    aspectRatio: Literal["16:9", "9:16", "21:9"] = "16:9"
     seed: int | None = None
     loras: list[LoraEntry] = Field(default_factory=list[LoraEntry])
 
@@ -484,12 +490,15 @@ class GenerateImageRequest(BaseModel):
     model_config = ConfigDict(strict=True)
 
     prompt: NonEmptyPrompt
+    model: ImageGenerationModelCheckpointID = "z-image-turbo"
     width: int = Field(default=1024, ge=16)
     height: int = Field(default=1024, ge=16)
     numSteps: int = Field(default=4, ge=1)
     numImages: int = Field(default=1, ge=1)
     imagePath: str | None = None
     strength: float = Field(default=0.6, ge=0.0, le=1.0)
+    # Z-Image text-to-image only: 0 = off, 1 = max composition variety across seeds.
+    variation: float = Field(default=0.0, ge=0.0, le=1.0)
 
 
 def _default_model_types() -> set[ModelCheckpointID]:
@@ -558,6 +567,7 @@ class TargetResolution(BaseModel):
     height: int = Field(gt=0)
 
 
+
 class RetakeRequest(BaseModel):
     model_config = ConfigDict(strict=True)
 
@@ -592,6 +602,42 @@ class ExtendRequest(BaseModel):
 ExtendResponse: TypeAlias = RetakeResponse
 
 
+class QwenMultiAngleGenerateRequest(BaseModel):
+    model_config = ConfigDict(strict=True)
+
+    image_data_url: str
+    # Compose mode: additional reference images (prop, location, wardrobe) fed to
+    # Qwen alongside the subject. image_data_url is the subject being re-angled;
+    # these are extra context the model composes from. Empty = classic single-image.
+    extra_image_data_urls: list[str] = Field(default_factory=list)
+    # Role of each extra ref, aligned index-for-index with extra_image_data_urls:
+    # "location" = a scene the subject is composited INTO, "prop" = an object the
+    # subject holds/uses. Empty (or length-mismatched) = classic compose, no
+    # explicit compositing instruction. Drives compose_prompt().
+    extra_image_roles: list[Literal["location", "prop"]] = []
+    azimuth_deg: float
+    elevation_deg: float
+    zoom: float
+    seed: int = 42
+    randomize_seed: bool = False
+    # Sampling recipe: "fast" = 4-step Lightning (~30s, softer skin), "balanced"
+    # = 8-step Lightning (retains more texture, ~2x fast), "quality" = 28-step
+    # base, no distillation LoRA (sharpest skin, slowest). See the GGUF pipeline.
+    quality_mode: Literal["fast", "balanced", "quality"] = "fast"
+    # Optional skin-realism LoRA stacked on top of the active mode; skin_weight
+    # scales it (recommended ~0.6-1.2) so it doesn't overpower the angles LoRA.
+    use_skin: bool = False
+    skin_weight: float = 1.0
+    extra_prompt: str = ""
+
+
+class QwenMultiAngleGenerateResponse(BaseModel):
+    status: Literal["complete"] = "complete"
+    image_data_url: str
+    prompt: str
+    seed: int
+
+
 ConditioningType: TypeAlias = Literal["canny", "depth"]
 
 # Generation can additionally run a user-supplied IC-LoRA against a pre-rendered

backend/app_factory.py

@@ -23,6 +23,7 @@
 from _routes.image_gen import router as image_gen_router
 from _routes.prompt_enhancement import router as prompt_enhancement_router
 from _routes.models import router as models_router
+from _routes.qwen_multiangle import router as qwen_multiangle_router
 from _routes.suggest_gap_prompt import router as suggest_gap_prompt_router
 from _routes.retake import router as retake_router
 from _routes.extend import router as extend_router
@@ -161,6 +162,7 @@ async def _route_generic_error_handler(request: Request, exc: Exception) -> JSON
     app.include_router(image_gen_router)
     app.include_router(suggest_gap_prompt_router)
     app.include_router(retake_router)
+    app.include_router(qwen_multiangle_router)
     app.include_router(extend_router)
     app.include_router(ic_lora_router)
     app.include_router(lora_catalog_router)