sail-sg/D-TRAK
Intriguing Properties of Data Attribution on Diffusion Models (ICLR 2024)
This project helps machine learning researchers and practitioners understand which specific training images most influence the outputs of a diffusion model. You input a trained diffusion model and a dataset, and it outputs visualizations showing the "proponent" (most supportive) and "opponent" (most unsupportive) training examples for any given generated image. This allows you to trace back and identify the data points that shaped a model's particular output.
No commits in the last 6 months.
Use this if you need to debug, explain, or improve the behavior of diffusion models by understanding the specific training data points that contribute to their generated outputs.
Not ideal if you are looking to attribute other types of machine learning models (e.g., discriminative classifiers) or if you are not working with diffusion models.
Stars
37
Forks
3
Language
Jupyter Notebook
License
MIT
Category
Last pushed
Jan 23, 2024
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/sail-sg/D-TRAK"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
quantgirluk/aleatory
📦 Python library for Stochastic Processes Simulation and Visualisation
blei-lab/treeffuser
Treeffuser is an easy-to-use package for probabilistic prediction and probabilistic regression...
TuftsBCB/RegDiffusion
Diffusion model for gene regulatory network inference.
yuanchenyang/smalldiffusion
Simple and readable code for training and sampling from diffusion models
chairc/Integrated-Design-Diffusion-Model
IDDM (Industrial, landscape, animate, latent diffusion), support LDM, DDPM, DDIM, PLMS, webui...