What the DAAM: Interpreting Stable Diffusion Using Cross Attention

Updated to support Diffusers 0.16.1!

I regularly update this codebase. Please submit an issue if you have any questions.

In our paper, we propose diffusion attentive attribution maps (DAAM), a cross attention-based approach for interpreting Stable Diffusion. Check out our demo: https://huggingface.co/spaces/tetrisd/Diffusion-Attentive-Attribution-Maps. See our documentation, hosted by GitHub pages, and our Colab notebook, updated for v0.1.0.

Getting Started

First, install PyTorch for your platform. Then, install DAAM with pip install daam, unless you want an editable version of the library, in which case do git clone https://github.com/castorini/daam && pip install -e daam. Finally, login using huggingface-cli login to get many stable diffusion models -- you'll need to get a token at HuggingFace.co.

Running the Website Demo

Simply run daam-demo in a shell and navigate to http://localhost:8080. The same demo as the one on HuggingFace Spaces will show up.

Using DAAM as a CLI Utility

DAAM comes with a simple generation script for people who want to quickly try it out. Try running

$ mkdir -p daam-test && cd daam-test
$ daam "A dog running across the field."
$ ls
a.heat_map.png    field.heat_map.png    generation.pt   output.png  seed.txt
dog.heat_map.png  running.heat_map.png  prompt.txt

Your current working directory will now contain the generated image as output.png and a DAAM map for every word, as well as some auxiliary data. You can see more options for daam by running daam -h.

Using DAAM as a Library

Import and use DAAM as follows:

from daam import trace, set_seed
from diffusers import StableDiffusionPipeline
from matplotlib import pyplot as plt
import torch


model_id = 'stabilityai/stable-diffusion-2-base'
device = 'cuda'

pipe = StableDiffusionPipeline.from_pretrained(model_id, use_auth_token=True)
pipe = pipe.to(device)

prompt = 'A dog runs across the field'
gen = set_seed(0)  # for reproducibility

with torch.cuda.amp.autocast(dtype=torch.float16), torch.no_grad():
    with trace(pipe) as tc:
        out = pipe(prompt, num_inference_steps=30, generator=gen)
        heat_map = tc.compute_global_heat_map()
        heat_map = heat_map.compute_word_heat_map('dog')
        heat_map.plot_overlay(out.images[0])
        plt.show()

You can also serialize and deserialize the DAAM maps pretty easily:

from daam import GenerationExperiment, trace

with trace(pipe) as tc:
    pipe('A dog and a cat')
    exp = tc.to_experiment('experiment-dir')
    exp.save()  # experiment-dir now contains all the data and heat maps

exp = GenerationExperiment.load('experiment-dir')  # load the experiment

We'll continue adding docs. In the meantime, check out the GenerationExperiment, GlobalHeatMap, and DiffusionHeatMapHooker classes, as well as the daam/run/*.py example scripts. You can download the COCO-Gen dataset from the paper at http://ralphtang.com/coco-gen.tar.gz. If clicking the link doesn't work on your browser, copy and paste it in a new tab, or use a CLI utility such as wget.

See Also

• DAAM-i2i, an extension of DAAM to image-to-image attribution.

• Furkan's video on easily getting started with DAAM.

• 1littlecoder's video for a code demonstration and Colab notebook of an older version of DAAM.

Citation

@inproceedings{tang2023daam,
    title = "What the {DAAM}: Interpreting Stable Diffusion Using Cross Attention",
    author = "Tang, Raphael  and
      Liu, Linqing  and
      Pandey, Akshat  and
      Jiang, Zhiying  and
      Yang, Gefei  and
      Kumar, Karun  and
      Stenetorp, Pontus  and
      Lin, Jimmy  and
      Ture, Ferhan",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    year = "2023",
    url = "https://aclanthology.org/2023.acl-long.310",
}