ComfyUI DAAM

ComfyUI custom nodes for Diffusion Attentive Attribution Maps (DAAM)

This extension enables visualization of cross-attention heatmaps within Stable Diffusion models, showing exactly which parts of the image correspond to specific words in the prompt.

Tested with SDXL and SD 1.5 models.

Example workflows are available in workflows.

This project was adapted from the SD Web UI implementation. Special thanks to @kousw for the original work!

Installation

Now Available on Comfy Registry and ComfyUI Manager!

Manual Install

Clone this repo into your ComfyUI custom_nodes directory:

git clone https://github.com/nisaruj/comfyui-daam.git

Then install the required packages

cd comfyui-daam
python -s -m pip install -r requirements.txt

Restart ComfyUI.

DAAM Nodes

CLIPTextEncodeWithTokens

Identical to CLIPTextEncode but also outputs the tokenized prompt required for the analysis.

KSamplerDAAM

A hooked version of KSampler. During sampling, it records attention maps for later analysis.

Outputs:

• latent — standard latent output
• heatmaps — raw heatmaps for input into the analyzer

DAAMAnalyzer

This node generates overlay heatmaps that show which parts of the image correspond to selected words in the prompt.

Inputs:

• clip — CLIP model used to encode the attention text
• tokens — from CLIPTextEncodeWithTokens
• heatmaps — from KSamplerDAAM
• images — the output images to overlay the heatmaps
• A text box for comma-separated words to generate heatmaps

Output:

• A batch of images with word-level heatmaps overlaid

Example Result

Prompt: A photo of corgi with a cowboy hat riding a skateboard in the park

Attention Words: corgi,skateboard,hat,park

Changelog

• 0.2.0
  • Added support for batched inputs
  • SD 1.5 models compatibility
  • Bug fixes and stability improvements
• 0.1.1
  • Initial Version with SDXL support