TIDE attributes diffusion outputs to training images in milliseconds
serrjoa · x · 2026-10-09
Adobe researchers (Shixuan Liu, Joan Serrà, et al.) published "Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution," introducing TID and TIDE to answer which training images influenced a generated image.
- Existing methods need costly per-sample gradients or query-specific optimization, and often attribute proxy losses rather than real generative behavior
- TID formulates attribution via a local score discrepancy measure, works with DDPM/EDM/flow matching without retraining, and uses Kronecker-factored curvature to avoid random projections
- TIDE distills TID into a forward-only student reproducing teacher rankings from internal activations
- On CIFAR-10, ArtBench-10, and MS-COCO, TID matches or beats SOTA while TIDE keeps most accuracy at 4-5 orders of magnitude lower per-query cost, attributing in milliseconds
More from Research
- Interactive Alignment paper uses evolutionary game theory to test long-run AI alignment — sebkrier · 2026-10-09
- Flow matching enables accurate prediction of partially occupied crystal structures — CatAstro_Piyush · 2026-10-09
- LightOn built an OCR-and-layout-detector annotation pipeline to train document grounding — IgorCarron · 2026-10-09
- MC-PhysBench: leading video world models break physics within a few frames — CatAstro_Piyush · 2026-10-09
- Nobel-related paper publishes its full referee reports—and even top LLMs can't summarize them — joshgans · 2026-10-09
- Scientist: As AI Tries Everything Useful, Basic Science Is a Better Bet Than Ever — OdedRechavi · 2026-10-09