Reproduce “Distributional Statistics Restore Training Data Auditability in One-step Distilled Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Muxing Li, Zesheng Ye, Sharon Li, Andy Song, Guangquan Zhang, Feng Liu
The authors introduce D-UTDD to audit whether a distilled image generator’s candidate dataset contains training data. It compares generated-image distributions with candidate data and a non-training reference using maximum mean discrepancy. In tests, detectors approached chance on DMD and Diff-Instruct students, while D-UTDD reached 92% accuracy at 30% training-data prevalence. This is dataset-level evidence, not proof of individual-image use or infringement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.