README.md

---
license: apache-2.0
datasets:
- ruchangyao/FlashVSR_Plus_Dataset
base_model:
- Wan-AI/Wan2.1-T2V-1.3B
pipeline_tag: video-to-video
tags:
- video-super-resolution
- video-restoration
- flashvsr
---

# FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching

<div align="center">

[**Project Page**](https://flashvsr.tech) | [**Paper**](https://arxiv.org/abs/2610.17552) | [**GitHub**](https://github.com/OpenImagingLab/FlashVSR_Plus) | [**Dataset**](https://huggingface.co/datasets/ruchangyao/FlashVSR_Plus_Dataset)

</div>

## Overview

**FlashVSR+** is an efficient one-step diffusion framework for high-quality video super-resolution, enabling **spatially untiled 4K inference on a single 24GB RTX 4090 GPU**.

FlashVSR+ introduces two key innovations:

- **Hybrid Attention:** Exploits spatial and temporal attention locality while retaining sparse global interactions, achieving up to **2× DiT acceleration** and **45% lower peak memory**.
- **Reward-Yield Distribution Matching Distillation (RY-DMD):** Enables preference alignment with black-box rewards without differentiating through the reward model, improving perceptual quality and generalizing to text-to-image generation.

For visual comparisons and additional results, please visit our [project page](https://flashvsr.tech).

## Released Models

We provide two FlashVSR+ variants:

| Model | Description |
|---|---|
| **FlashVSR+ Turbo** | Faster inference, recommended for efficient high-resolution video super-resolution. |
| **FlashVSR+ Base** | Higher-quality restoration, recommended when reconstruction quality is the priority. |

Both models support one-step inference and share the same decoder.

## Model Files

The released checkpoints are organized as follows:

```text
ckpts/
├── turbo/
│   ├── dit.safetensors
│   └── lq_proj_in.safetensors
├── base/
│   ├── dit.safetensors
│   └── lq_proj_in.safetensors
├── TCDecoder.ckpt
└── Wan2.1_VAE.pth
```

- `dit.safetensors`: Main diffusion transformer weights.
- `lq_proj_in.safetensors`: Low-quality video conditioning weights.
- `TCDecoder.ckpt`: Shared decoder used by the default inference pipeline.
- `Wan2.1_VAE.pth`: Optional VAE decoder weights.

## Citation

If you find FlashVSR+ useful, please cite:

```bibtex
@misc{yao2026flashvsrefficientvideosuperresolution,
  title={FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching},
  author={Ruchang Yao and Shi Guo and Junhao Zhuang and Mingde Yao and Xin Cai and Jiahao Zhan and Bin Fu and Xinyuan Chen and Yu Qiao and Tianfan Xue},
  year={2026},
  eprint={2610.17552},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.17552}
}
```

## Acknowledgements

FlashVSR+ builds upon the original [FlashVSR](https://arxiv.org/abs/2510.12747). We thank the authors of the open-source projects and pretrained models that support this work.