---
license: mit
title: Wan
sdk: docker
emoji: 🏆
colorFrom: blue
colorTo: blue
pinned: true
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/690c81642c226d46d51dc1a5/EhLyLqcHHk_Q_gb7xR1HF.jpeg
short_description: Wan
---
# AI Video Swarm

8× L40S GPU Video Generation Platform on Hugging Face Spaces

## Hardware
- **8× NVIDIA L40S** (48GB VRAM each = 384GB total)
- **192 vCPU**
- **1534 GB RAM**

## File Structure
```
.
├── app.py                      # FastAPI main app
├── index.html                  # Main UI (single file, RTL Arabic)
├── pass.html                   # Auth gateway
├── logs.html                   # Real-time logs dashboard
├── requirements.txt            # Python dependencies
├── Dockerfile                  # HF Spaces compatible
├── README.md                   # This file
├── agents/
│   ├── __init__.py
│   ├── orchestrator.py         # Master coordinator
│   ├── prompt_engineer.py      # LLM prompt enhancement
│   ├── image_generator.py      # FLUX.1-dev keyframe
│   ├── video_generator.py      # Wan 2.6 + LTX-2.3
│   ├── face_restorer.py        # GFPGAN (placeholder)
│   ├── frame_interpolator.py   # RIFE (placeholder)
│   ├── video_upscaler.py       # Real-ESRGAN (placeholder)
│   └── log_monitor.py          # Logging agent
├── config/
│   ├── __init__.py
│   ├── settings.py             # App config
│   ├── models_config.py        # Model params
│   └── gpu_allocation.py       # GPU mapping
└── core/
    ├── __init__.py
    ├── auth.py                 # JWT auth
    ├── r2_client.py            # Cloudflare R2
    └── websocket_manager.py    # WS connections
```

## Environment Variables (Secrets)
Set these in HF Space Settings > Secrets:
- `HF_TOKEN` — HuggingFace access token
- `PASS_KEY` — Password for login
- `R2_ACCOUNT_ID` — Cloudflare R2 account ID
- `R2_BUCKET_NAME` — R2 bucket name
- `R2_ACCESS_KEY_ID` — R2 access key
- `R2_SECRET_ACCESS_KEY` — R2 secret key
- `R2_PUBLIC_DOMAIN` — R2 public domain

## Models Used
| Agent | Model | GPU | VRAM |
|-------|-------|-----|------|
| Video Primary | Wan 2.6 14B (FP8) | GPU-0 | ~45GB |
| Video Backup | LTX-2.3 | GPU-1 | ~20GB |
| Image | FLUX.1-dev (FP8) | GPU-5 | ~16GB |
| Upscale | Real-ESRGAN | GPU-4 | ~8GB |
| Face Restore | GFPGAN | GPU-2 | ~8GB |
| Interpolation | RIFE | GPU-3 | ~6GB |
| Audio | F5-TTS | GPU-6 | ~6GB |
| Orchestrator | Qwen 2.5 72B | GPU-7 | ~40GB |

## Notes
- PyTorch wheels include CUDA 12.4 — no need for nvidia/cuda base image
- All models load at startup (no cold start)
- FP8 inference on L40S for 2× speed
- NVENC hardware encoding for final output