PerturbedAttentionGuidance - ComfyUI Built-in Node Documentation

The PerturbedAttentionGuidance node applies perturbed attention guidance to a diffusion model to improve generation quality. During sampling it makes an extra prediction in which the middle block’s self-attention is replaced by a simplified version that passes the value projections straight through, then adds the scaled difference between the normal conditional prediction and this perturbed prediction to the denoised result. Setting scale to 0 disables the effect entirely.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The diffusion model to apply perturbed attention guidance to	MODEL	Yes	-
`scale`	The strength of the perturbed attention guidance effect (default: 3.0). When set to 0, the node has no effect and returns the original denoised result unchanged.	FLOAT	Yes	0.0 - 100.0 (step: 0.01)

Outputs

Output Name	Description	Data Type
`model`	The modified model with the perturbed attention guidance patch attached to its sampling process	MODEL

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 1cf824486ae695a9e563c70a4798aaf4c9c067ae3b53172c9767e3c5093d0096

​Inputs

​Outputs

Inputs

Outputs