# FlashVSR Model Mirror (AEmotionStudio)

Mirror of [FlashVSR v1.1](https://github.com/OpenImagingLab/FlashVSR) model weights for [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA).

## Model Files

| File | Description | Size |
|---|---|---|
| `FlashVSR1_1.safetensors` | DiT v1.1 weights (one-step diffusion backbone) | ~5.7 GB |
| `Wan2.1_VAE.safetensors` | WAV2.1 VAE decoder | ~254 MB |
| `LQ_proj_in.safetensors` | Low-quality 4× projection encoder | ~576 MB |
| `TCDecoder.safetensors` | Tiny Conditional Decoder (for Tiny/Tiny-Long modes) | ~181 MB |
| `Prompt.safetensors` | Cross-attention conditioning embeddings | ~4 MB |

## Variants

- **FlashVSR Full** — Best quality, full VAE decode (~12-16 GB VRAM)
- **FlashVSR Tiny** — Fast mode with TCDecoder (~8-12 GB VRAM)
- **FlashVSR Tiny Long** — Streaming for long videos, lowest VRAM (~8-12 GB)

All variants share the same weight files. The variant determines which pipeline and decoder components are used at inference time.

## Usage

These models are automatically downloaded by [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA) when selecting any `flashvsr_*` model from the `upscale_model` dropdown.

Manual installation:
```
ComfyUI/models/FlashVSR/
├── FlashVSR1_1.safetensors
├── Wan2.1_VAE.safetensors
├── LQ_proj_in.safetensors
├── TCDecoder.safetensors
└── Prompt.safetensors
```

## Attribution

- **Original model**: [FlashVSR (CVPR 2026)](https://github.com/OpenImagingLab/FlashVSR) by Junhao Zhuang et al.
- **Original weights**: [JunhaoZhuang/FlashVSR](https://huggingface.co/JunhaoZhuang/FlashVSR)
- **ComfyUI adaptation (BSA-free)**: [1038lab/ComfyUI-FlashVSR](https://github.com/1038lab/ComfyUI-FlashVSR)
- **ComfyUI integration weights**: [1038lab/FlashVSR](https://huggingface.co/1038lab/FlashVSR)

## License

GPL-3.0 — See [LICENSE](LICENSE) for full text.

The FlashVSR model weights are released under the GNU General Public License v3.0.
All usage must comply with the original license terms from the FlashVSR authors.