README.md

metadata

title: Workshop
emoji: 🛠️
colorFrom: yellow
colorTo: red
sdk: docker
pinned: false
license: mit
short_description: LoRA pipeline configurator for Wan 2.2 I2V

Workshop

Configure a Wan 2.2 I2V Lightning 14B base pipeline or the Qwen-Image-Edit-2511 LoRA Studio (multi-image examples, T2I edit), attach LoRAs from your reference library (00000tt/wan22-loras), label example image + prompt pairs, and publish a private self-contained Gradio Space.

Templates

1. Wan 2.2 I2V Lightning 14B (AoT FP8) — video generation; LoRA pulled from the warehouse on the first generate (cached after). templates/wan22_i2v/.
2. Qwen-Image-Edit-2511 LoRAs Fast — image editing; the proven Qwen-Image-Edit-2511-LoRAs-Fast recipe (Rapid-AIO transformer, FA-attn3, lazy LoRA loading, 20+ LoRAs). Examples support multiple images per preset (lighting/style transfer). templates/qwen_image_edit_2511/.

Workflow

1. Vault — the Space's HF_TOKEN secret auto-loads. A manual token is only a fallback.
2. Model & pipeline — pick the base pipeline (Wan 2.2 I2V Lightning 14B, FP8 + int8 text-encoder, 6 steps default).
3. LoRA reference library — checkbox the .safetensors you want attached; drag new ones into the Reference library parts panel to upload them to the warehouse.
4. Generation settings — steps, guidance, duration stay as defaults in the published Space.
5. Examples — label image + prompt pairs; they are baked into the published Space's presets.
6. Publish — creates a Space under 00000tt, private by default. Flip visibility to public only when you want it visible.

Assistant

The Workshop Assistant talks through the Hugging Face Inference API using the same HF_TOKEN secret. It can list the warehouse, set pipeline config, publish private Spaces, flip visibility — and research the Hub: ask it to find similar models and it will search Spaces, rank them by likes, fetch the target app.py, extract that workflow (pipeline, base model, steps, guidance, fps, frames, LoRA use, quantization), and diff it against your configured pipeline so you can decide whether to match it or read their source. It also carries a gold-standard rubric distilled from your proven reference Space (wan2-2-fp8da-aoti-preview-nsfw) plus diffusers best practice — say "diagnose mine" and it grades every Space you've published as Perfect / Solid / Needs polish / Unreliable with a per-check pass/fail list (base resolves, right pipeline, dtype, quantization with fallback, VRAM cleanup, resolution, seeds, frame math, LoRA token flow, valid examples, negative-prompt, controls, @spaces.GPU, no leaked secrets) and proposes add-ons. All driven by small JSON tool calls at the end of its replies, parsed and executed by the backend; a rule-based fallback answers when the LLM is unavailable.

Develop locally

npm install && npm run dev   # frontend dev server (proxies /api to :7860)
uvicorn server:app --port 7860