castorini/daam
A cross-attention visualization technique for interpreting Stable Diffusion image generation.
Not currently ranked — collecting fresh signals.
star history
DAAM (Diffusion Attentive Attribution Maps) generates attention-based heatmaps that reveal how Stable Diffusion models respond to different parts of a text prompt when generating images. The tool hooks into the cross-attention layers of diffusion models to produce attribution maps that highlight which word tokens most influence different regions of the generated output. It supports the standard Stable Diffusion models as well as SDXL, and provides both a Python library for integration and a CLI/demo interface.
Frequently asked
- What is castorini/daam?
- A cross-attention visualization technique for interpreting Stable Diffusion image generation.
- Is daam open source?
- Yes — castorini/daam is open source, released under the MIT license.
- What language is daam written in?
- castorini/daam is primarily written in Jupyter Notebook.
- How popular is daam?
- castorini/daam has 801 stars on GitHub.
- Where can I find daam?
- castorini/daam is on GitHub at https://github.com/castorini/daam.