Practical overview
When it fits
Use DAAM when you already have a Stable Diffusion generation workflow and need word-level attribution maps to examine prompt-image relationships or preserve results for later inspection.
Who it is for
- Developers building or adapting Stable Diffusion generation workflows.
- Researchers and practitioners inspecting how prompt terms relate to generated image regions.
Documented uses
- Generate an image with a separate attribution heat map for each prompt word.
- Overlay a selected word's heat map on a generated image in Python.
- Save and reload generation experiments and their attribution data.
- Inspect Stable Diffusion XL outputs through the documented CLI or Diffusers workflow.
Requirements & limits
The repository is public and lists an MIT license. The documented installation uses PyTorch and pip; access to many Stable Diffusion models requires a Hugging Face token and CLI login.
- DAAM is an interpretation tool that traces a generation pipeline; the documented examples use a Diffusers pipeline for image generation.
- The CLI saves a heat map per prompt word, while the Python example shows computing and overlaying a selected word map.
- The README identifies support for SDXL and Diffusers 0.21.1.
- The documented SDXL Python example uses CUDA, but the CLI example does not state a hardware requirement.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Generate an image with word-level attribution maps
Use the DAAM command-line utility to generate a Stable Diffusion image and save its prompt-word heat maps.
Before you start
- Install PyTorch for the platform.
- Install DAAM with pip.
- Log in to Hugging Face with a token to access many Stable Diffusion models.
Steps
Create and enter a working directory.
Run
daam "A dog running across the field."from the shell.Review the generated
output.png, per-word.heat_map.pngfiles, and auxiliary data in the working directory.
Expected outcome
The working directory contains the generated image, a heat map for each prompt word, and auxiliary generation files.
Guide limitations
- The README notes that access to many Stable Diffusion models requires Hugging Face login and a token.
- The README does not specify hardware requirements for this CLI example.
Save and reload a traced generation experiment
Preserve DAAM attribution maps and generation data from a Diffusers pipeline, then load the saved experiment.
Before you start
- Install PyTorch and DAAM.
- Have a Diffusers pipeline and a Stable Diffusion model available; access to many models requires Hugging Face login and a token.
Steps
Trace a pipeline with DAAM while running a prompt through it.
Convert the trace to an experiment with
tc.to_experiment('experiment-dir').Save the experiment with
exp.save().Load it later with
GenerationExperiment.load('experiment-dir').
Expected outcome
The experiment directory contains saved data and heat maps that can be loaded as a `GenerationExperiment`.
Guide limitations
- The README's example assumes a pipeline has already been set up.
- The illustrated model-loading example uses CUDA and a specific SDXL configuration; it does not establish requirements for every model or device.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- castorini/daam · Official homepage title · retained in its source language