ComfyUI-RadialAttn

This repo is RadialAttention ported to ComfyUI native workflows. If you're using kijai's ComfyUI-WanVideoWrapper rather than native workflows, then you can use their WanVideoSetRadialAttention node rather than this repo, and you still need to install the pip packages below.

This supports Wan 2.1 14B, Wan 2.2 14B, Wan 2.2 5B, both T2V and I2V.

Installation

1. Install SpargeAttention
2. git clone this repo to your ComfyUI/custom_nodes/

It's also recommended to install SageAttention, and add --use-sage-attention when starting ComfyUI. When RadialAttention is not applicable, SageAttention will be used.

Usage

Just connect your model to the PatchRadialAttn node. There's an example workflow for Wan 2.2 14B I2V + GGUF + LightX2V LoRA + RadialAttention + torch.compile.

It's believed that skipping RadialAttention on the first layer (dense_block = 1) and the first time step (dense_timestep = 1) improves the quality.

RadialAttention requires specific video sizes and lengths:

• The 'number of video tokens' must be divisible by 128, see video_token_num for details
• For Wan 2.1 and 2.2 14B, this number is computed by width/16 * height/16 * (length+3)/4
• For Wan 2.2 5B, this number is computed by width/32 * height/32 * (length+3)/4

(A misunderstanding is that the width and the height must be divisible by 128, but that's actually not the case.)

Don't blindly use torch.compile. To start with, you can disable the TorchCompileModel node and run the workflow. Only when you're sure that the workflow runs but it's not fast enough, then you can try to enable TorchCompileModel. There are reports that torch.compile is slower in PyTorch 2.8 .