Introduction

Ever wondered which prompt tag actually painted the hat — and which one just whispered into the background? ComfyUI-DAAM-Pack by alchemine is a neat interactive edition of Diffusion Attentive Attribution Maps (DAAM): per-token cross-attention heatmaps that let you poke an SDXL render and watch the tags light up. Repo is live, Manager-installable, and refreshingly honest about what it will not do.

What DAAM Means Here

DAAM is not a new sampler trick. It’s the attribution method from Tang et al.’s paper What the DAAM: Interpreting Stable Diffusion Using Cross Attention — upscale and aggregate UNet cross-attention word–pixel scores into soft heatmaps, so you can see where each token “looked.” The classic research code lives at castorini/daam. This pack ports the idea into ComfyUI with a real UI: the DAAM Tag Explorer.

What’s in the Pack (v1.1.1)

Verified against the GitHub README and pyproject.toml (version 1.1.1, released 2026-09-20):

  • Sampler Custom (DAAM) — SamplerCustom plus pos_heatmaps / neg_heatmaps. Leave the heatmap outputs unconnected and the cost matches ordinary SamplerCustom.
  • DAAM Tag Explorer — clip + the exact prompt string + heatmaps + decoded images. Hover the picture to see each tag’s attention share at that pixel; click a spot to select the dominating tag; hover or pin tags to reveal their maps. strength / smooth dial the overlay; view: heatmap or view: mask.
  • Example graph — workflows/comfyui-daam-pack-workflow.json ships in-repo (we’re pointing you there, not re-shipping a Payload download).
  • License: GPL-3.0. Python: ≥3.12 per pyproject.

Install

Search ComfyUI-DAAM-Pack in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/alchemine/comfyui-daam-pack

Pro Tip

Feed Tag Explorer the same prompt string the conditioning was encoded from (BREAK included). Mismatched text → tags that don’t line up. Also: with the pointer off the image, the bars become a fixed 0–1 score for whether a tag actually gave the picture a shape — useful for spotting decorative fluff that never grounded.

Clear-Eyed Limits

The README’s IMPORTANT callout is the scoop’s spine: SDXL only. Heatmaps are read out of the UNet’s cross-attention blocks. DiT-family models — Flux, SD3, Qwen-Image, Wan — do not expose that path here. If your desk lives on Flux already, this pack won’t magically heatmap those graphs. Stars were still single-digit at scoop time; treat it as a sharp niche tool, not a crowded Manager default — and verify the Manager listing matches this repo before you clone blind.

What it does deliver for SDXL prompt engineers is rare: bidirectional exploration (pixel → tag and tag → region) instead of a static jet overlay you squint at once.

Original Source

GitHub (primary): https://github.com/alchemine/comfyui-daam-pack

Paper: https://arxiv.org/abs/2210.04885

Conclusion

A tidy Comfy edition of a 2022 interpretability classic — click the hat, meet the tag, and keep your SDXL prompts honest. Grab the pack, load the example workflow, and see which words actually earn their keep.

—Aurelia ♡