ComfyUI-RAVE_ATTN

ComfyUI-RAVE_ATTN
★ 13

时序注意力RAVEComfyUI节点UNet自注意力
在ComfyUI中实现RAVE时序注意力:在UNet的Self-Attention中应用RAVE技术,不拼接图像,保留对其他时序和样式机制的兼容性。
💡 在ComfyUI流水线中为动画/视频增强不破坏UNet的时序注意力。
🍴 3 Forks💻 Python🔄 2024-05-22
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/8f9eee5e2cdb
📄 README

ComfyUI-RAVE Attention

ComfyUI nodes to use RAVE attention as a temporal attention mechanism.

This differs from other implementations in that it does not concatenate the images together, but within the UNet’s Self-Attention mechanism performs the RAVE technique.

By not altering the images/latents throughout the UNet, this method does not affect other temporal techniques, style mechanisms, or other UNet modifications.

For example, it can be combined with AnimateDiff, ModelScope/ZeroScope, or FLATTEN.

Installation

This repository does not require any additional packages. Just git clone this into your custom_nodes/ directory.

Examples

See the example_workflows directory for example workflows.

Acknowledgements

  • Ozgur Kara, Bariscan Kurtkaya, Hidir Yesiltepe, James M. Rehg, Pinar Yanardag for researching and open sourcing their implementation of RAVE