ComfyUI-SolAttn_triton

<h1 align="center">ComfyUI-SolAttn</h1> <h4 align="center"> Experimental Triton implementation of Sol-Attn for ComfyUI </h4> <p align="center"> <a href="https://arxiv.org/abs/2607.24027"><img src="https://img.shields.io/badge/📄_Paper-arXiv-b31b1b?style=flat-square" alt="Paper"/></a> <a href="https://github.com/NVlabs/Sana/tree/sol-engine/techniques/sparse_backends/sol_attn"><img src="https://img.shields.io/badge/💻_Code-Sol--Attn-76b900?style=flat-square" alt="Code"/></a> <a href="https://nvlabs.github.io/Sana/Sol-Attn/"><img src="https://img.shields.io/badge/🌐_Project-Page-blue?style=flat-square" alt="Project Page"/></a> </p>

---

Overview

Sol-Attn is a training-free sparse attention method for accelerating image and video generation. This community extension integrates a Triton implementation of Sol-Attn into ComfyUI.

[!NOTE] This project is a work in progress. It has currently been tested on RTX 4090 and RTX 5090 GPUs with MiniMax H3.

Usage notes

Triton kernels are compiled on first use, so the first run will be slower.

Use start_percent, end_percent, and tau to balance generation quality and speed.

Examples

Test output

https://github.com/user-attachments/assets/8d9ed820-0417-4d68-9d1c-5199534bed3b

SageAttention vs. Sol-Attn

<table> <tr> <td align="center"><b>SageAttention</b></td> <td align="center"><b>Sol-Attn</b></td> </tr> <tr> <td width="50%"> <video src="https://github.com/user-attachments/assets/27f201ea-6bfc-4f43-826c-51809eed9d15" controls muted loop></video> </td> <td width="50%"> <video src="https://github.com/user-attachments/assets/73f63d14-2166-4f62-b098-e817ec1d7704" controls muted loop></video> </td> </tr> </table> <img width="482" height="500" alt="Sol-Attn example result" src="https://github.com/user-attachments/assets/27ae9886-aa3e-4470-a507-3a7c52b24be5" />

Citation

If you find Sol-Attn useful in your work, please cite the paper:

@article{li2026solattn,
  title={Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification},
  author={Li, Haopeng and Li, Yitong and Chen, Junsong and Ye, Tian and Liu, Haozhe and Yu, Jincheng and Wang, Duomin and Zhang, Ruihua and Xie, Zeke and Xie, Enze and Han, Song},
  journal={arXiv preprint arXiv:2607.24027},
  year={2026}
}