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

Project Page | Paper | GitHub | Dataset

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.

Released Models

We provide two FlashVSR+ variants:

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

Model Files

The released checkpoints are organized as follows:

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:

@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. We thank the authors of the open-source projects and pretrained models that support this work.