ComfyUI Node
Attention Maps per word
One heatmap per prompt word: where in the picture that word was attended to, averaged over heads, layers and steps. Outputs a sheet of the picture with one overlay per word, the maps as a MASK batch in word order (each stretched between its own floor and peak), and the data as JSON.
Attention Maps per word
- attention
- clip
- image
- sheet
- maps
- json
◄texta red hat on a wooden table►
◄skip_punctuationtrue►
CategoryParametric/Sampling
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| attention | PSN_ATTENTION | Attention. | |
| clip | CLIP | The text encoder. | |
| text | STRING | a red hat on a wooden table | The input text. Default a red hat on a wooden table. |
| image | IMAGE | The input image. | |
| skip_punctuation | BOOLEAN | true | Skip punctuation. Default True. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| sheet | IMAGE | — |
| maps | MASK | — |
| json | STRING | — |