Concept Attention Visualizer

What it is

The display half of the pack. Concept Attention Visualizer takes the CONCEPT_MAPS bundle from ConceptAttentionNode plus the original image, blends the attention maps on top in color, and hands back a single IMAGE you can look at or save. If the pack's other nodes are the analysis, this is the "show me" node - the thing you wire up last, after you've got maps worth looking at.

Inputs and outputs

The whole thing in one line:

• concept_maps (CONCEPT_MAPS) - from ConceptAttentionNode.
• image (IMAGE) - the original, for the overlay base.
• visualized_image (IMAGE) - the composited result, straight into a PreviewImage or SaveImage.

Two inputs, one output. There are no knobs to tune here, which is either a blessing or the point where you start wanting more.

Honest notes

The README mentions an overlay_alpha transparency control. The shipped node doesn't have one. Reading the source, the blend strength is hardcoded - a fixed weight per concept, cycling through a small color palette as concepts pile up. So no tuning: what you get is what you get. The visualized_image output from ConceptAttentionNode is a similar single-image path if you want the pack's own default rendering, and if you want real per-concept control, you can pull individual saliency images out of ConceptSaliencyMapNode and composite them yourself with Blend nodes. That's strictly more flexible - this node just saves you the graph.

When you'd reach for it

As a sanity check on what the model actually found before you burn tokens on masks, or as a quick visual to drop into a write-up. It's the friendliest node in the pack, which is also why it's the least interesting one - you'd never reach for it on its own, only as the tail of a concept-attention workflow.

How to install

Same pack as the rest. In ComfyUI Manager, search "Concept-Diffusion", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Junst/ComfyUI-Concept-Diffusion
cd ComfyUI-Concept-Diffusion
pip install -r requirements.txt

Restart ComfyUI. No model downloads - it only consumes the concept maps already in your graph. And because it's part of a young, single-commit pack, keep the ComfyUI console open on first run; if the maps are empty, this node will just quietly echo your image back, which is the pack's universal fallback and your cue to check upstream.