Multi-mode Gradio Studio for the Alibaba Wan video diffusion family (Wan 2.1 + Wan 2.2), targeting HF ZeroGPU (Blackwell) and local MPS (Apple Silicon).
See RESEARCH.md for the full architecture brief, model inventory, mode deep-dive, ZeroGPU integration recipe, Lightning LoRA strategy, UX wireframes, and dependency matrix.
Sidebar + header + per-mode tabs render. No actual generation yet. Wireframes for every screen are in wireframes/ (open wireframes/index.html for the gallery).
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.pyOpen http://localhost:7860 in your browser.
wan-studio/
├── RESEARCH.md # full architecture brief (1100+ lines)
├── app.py # Gradio entry point
├── requirements.txt # pinned deps (torch 2.8+, diffusers 0.38+, etc.)
├── pipelines/
│ ├── registry.py # source-of-truth (gen, mode, checkpoint) catalog
│ ├── shared.py # text-encoder / VAE / image-encoder shared loaders
│ ├── preset.py # Fast / Quality preset resolver with graceful fallback
│ └── __init__.py
├── ui/
│ ├── header.py # generation dropdown + Fast/Quality radio + nav icons
│ ├── sidebar.py # left mode picker
│ ├── tabs.py # per-mode panels (T2V, I2V, FLF2V, VACE, S2V, Animate, ...)
│ └── __init__.py
├── utils/
│ ├── backend.py # device + dtype detection + ZeroGPU awareness
│ └── __init__.py
├── tests/ # pytest stubs (Phase 1+)
├── assets/ # any static files served at /file
├── raw/ # research artifacts (per-topic deep dives + reference Spaces)
└── wireframes/ # 8 PNG mockups + index.html gallery
See RESEARCH.md §11 — 9 phases, ~3-4 weeks total, MVP in 7-9 days.
Current: Phase 0 — Scaffold + UI shell. Next: Phase 1 — T2V + I2V on Wan 2.1 14B end-to-end.