daam

Why we included this project

Stable Diffusion is something of a black box: you type a prompt, get an image, and usually can't say why a particular region of the picture came out the way it did. DAAM opens that box. It reads the cross-attention layers of a diffusion pipeline and, for each word in your prompt, produces a heat map showing where that concept actually landed in the generated image. That turns prompt debugging from guesswork into something you can look at directly, and it's a reasonable way to check whether a model is attending to the right regions before you trust its output. You can use it as a small CLI that drops per-word heat maps next to your generated image, or as a Python library wrapped around any Diffusers pipeline, with SDXL support included. For researchers and teams doing interpretability work on generative vision models, it's a solid, documented starting point.