MIT researchers demonstrate exact training-data deletion method for diffusion models

MIT CSAIL researchers reported in Nature Communications that a diffusion-ensemble architecture can remove the influence of individual training examples without retraining a model from scratch. Across 24 ensembles trained on datasets ranging from 256 to more than 160,000 images, they found that the maximum effect of removing one example declined as training sets grew. The team also retrained 1,282 small-scale models to test the finding. The work examines diffusion models, not large language models.