ComfyUI_FlashVSR

Update

• use new loop / 换一个loop切分超长视频
• 支持flashVSR-Plus 模型，plus模式 请在菜单手动选triton-flashvsr_attn，否则还会是 自动选择Block-Sparse-Attention ，support flashVSR-Plus model，plus use triton attention，or auto will choice Block-Sparse-Attention
• | plus_temporal | 4 | 每 latent 帧的输入帧数；大运动视频建议 2 或 1（更慢） |

git clone https://github.com/smthemex/ComfyUI_FlashVSR

2.requirements ----

pip install -r requirements.txt

• 1.optional requirements ( for flashvsr_attn)

pip install -e operators/flashvsr_attn --no-build-isolation --no-deps

• 2.optional requirements ( for Block-Sparse-Attention)

git clone https://github.com/mit-han-lab/Block-Sparse-Attention  # if 5090 or newest cuda
git clone https://github.com/smthemex/Block-Sparse-Attention # 无须梯子强制编译
git clone https://github.com/lihaoyun6/Block-Sparse-Attention # 须梯子
cd Block-Sparse-Attention
pip install packaging
pip install ninja
python setup.py install

3.checkpoints ----

• 3.1.1 FlashVSRv1.0 all checkpoints 所有模型，vae 用常规的wan2.1
• 3.1.2 FlashVSRv1.1 all checkpoints 所有模型，vae 用常规的wan2.1
• 3.1.3 FlashVSRv-plus all checkpoints 所有模型，vae 用常规的wan2.1
• 3.2.1 1.0/1.1 emb posi_prompt.pth 4M而已
• 3.2.2 plus emb posi_prompt.pth 4M而已
• 3.3 lightvaew2_1.pth and diffusion_pytorch_model.safetensors

├── ComfyUI/models/FlashVSR
|     ├── LQ_proj_in.ckpt # v1.1 or v1.0
|     ├── TCDecoder.ckpt
|     ├── diffusion_pytorch_model_streaming_dmd.safetensors #v1.1 or v1.0
|     ├── posi_prompt.pth
|     ├──lq_proj_in_plus_turbo.safetensors   #rename turbo or base
|     ├── base_dit_plus.safetensors #rename base or turbo
├── ComfyUI/models/vae
|        ├──Wan2.1_VAE.pth # or safetensors  optional/可选
|        ├──lightvaew2_1.pth  #32.2M  or taew2_1.pth,lighttaew2_1.pth optional/可选
|        ├──Wan2.1_VAE_upscale2x_imageonly_real_v1_diff.safetensors  # rename from diffusion_pytorch_model.safetensors optional/可选

Example

Previous

• 如果觉得项目有用，请给官方项目FlashVSR and FlashVSR_Plus打星；
• Block-Sparse-Attention 源码库已更新，支持5090系显卡及更新的设备。注意，需要cuda>=12.8
• 编译Block-Sparse-Attention window的轮子 可以使用 smthemex 强制编译版 或者 lihaoyun6 要联网 两个fork来
• Block-Sparse-Attention 正确安装且能调用才是方法的完全体，当前的函数实现会更容易OOM,但是Block-Sparse-Attention轮子实在不好找，目前只有CU128 toch2.7的，我提供的（cu128，torch2.8，py311单体,torch2.12+cu130 py311 py313）或者自己编译
• tile关闭质量更高，需要VRam更高，corlor fix对于非模糊图片可以试试。
• Choice vae infer full mode ，encoder infer tiny mode 选择vae跑full模式 效果最好，tiny则是速度，数据集基于4倍训练，所以1 scale是不推荐的；
• if you Like it ， star the official project： FlashVSR
• Good news, the mit-han-lab/Block-Sparse-Attention library has been updated and now supports 5090 and updated graphics cards ,need cuda>=12.8;
• The wheel for compiling the Block Spark Attention window can be used smthemex 或者 lihaoyun6Two forks
• The correct installation and ability to call Block Spark Attention are the complete set of methods. The current function implementation will be easier to OOM, but the Block Spark Attention wheel is really hard to find. Currently, only [CU128 toch2.7]（ https://github.com/lihaoyun6/ComfyUI-WanVideoWrapper ）Yes, I provided cu128，torch2.8，py311,torch2.12+cu130 py311 py313Or compile it yourself
• Tile shutdown has higher quality and requires higher VRam. For non blurry images, corlor fix can be tried.
• Choose VAE in full mode, encoder in tiny mode. Choosing VAE to run in full mode yields the best results, while tiny mode is for speed. The dataset is based on 4x training, so 1 scale is not recommended;

Citation

@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}, 
}
@article{zhuang2025flashvsr,
  title={FlashVSR: Towards Real-Time Diffusion-Based Streaming Video Super-Resolution},
  author={Zhuang, Junhao and Guo, Shi and Cai, Xin and Li, Xiaohui and Liu, Yihao and Yuan, Chun and Xue, Tianfan},
  journal={arXiv preprint arXiv:2510.12747},
  year={2025}
}

@misc{lightx2v,
 author = {LightX2V Contributors},
 title = {LightX2V: Light Video Generation Inference Framework},
 year = {2025},
 publisher = {GitHub},
 journal = {GitHub repository},
 howpublished = {\url{https://github.com/ModelTC/lightx2v}},
}