TomePatchModel - ComfyUI Built-in Node Documentation

TomePatchModel applies Token Merging (ToMe) to a diffusion model to reduce computational cost during inference.

TomePatchModel applies Token Merging (ToMe) to a diffusion model to reduce computational cost during inference. It works by merging similar tokens inside the model’s attention mechanism, so the model processes fewer tokens while keeping output quality largely intact.

Inputs

Parameter	Description	Data Type	Required	Range
model	The diffusion model to apply token merging to	MODEL	Yes	-
ratio	The proportion of tokens to merge (default: 0.3). Higher values merge more tokens, which can give greater speedup but potentially lower quality.	FLOAT	Yes	0.0 - 1.0

Note: If the number of tokens in an attention block is small enough that no downsampling is needed, the merging functions are replaced with no-ops, and the model runs unchanged for that block.

Outputs

Output Name	Description	Data Type
model	The modified model with token merging applied	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): 1202c0df17f357440cd156fa0920f70c18a318e32c41dc04cecff11613f0072f