AUTOMATIC1111
[Feature Request]: Add "Diffusion attentive attribution maps" to help understand how prompts affects image generation
Is there an existing issue for this?
- [X] I have searched the existing issues and checked the recent builds/commits
What would your feature do ?
https://github.com/castorini/daam would allow users to understand how different parts of prompt affect the generated image. Would require a custom implementation since of the interface the method in the repo doesn't seem to scale well to long prompts at all.
Example of what it can do:
Proposed workflow
Probably implemented as a custom extension that can hook up to txt2img (and img2img too) and add some kind of selector/dropdown to choose what parts of the prompt you want to see a heatmap for.
Additional information
No response
That looks really useful when trying to understand impact of promt keywords. If someone could make it work within the UI that would be amazing.
Related: https://github.com/benkyoujouzu/stable-diffusion-webui-visualize-cross-attention-extension
Related: https://github.com/benkyoujouzu/stable-diffusion-webui-visualize-cross-attention-extension
This sounds good but i'd prefer the daam style heatmap instead of this one
