ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features
Alec Helbling¹, Tuna Meral², Ben Hoover¹³, Pinar Yanardag², Duen Horng (Polo) Chau¹
¹Georgia Tech · ²Virginia Tech · ³IBM Research
We introduce ConceptAttention, an approach to interpreting the intermediate representations of diffusion transformers. The user just gives a list of textual concepts and ConceptAttention will produce a set of saliency maps depicting the location and intensity of these concepts in generated images. Check out our paper: here.
0 10000
0 18
0 4
Examples
| Enter your prompt | Enter a list of concepts (comma-separated) | Seed |
|---|