Training data attribution in diffusion models via mirrored unlearning and noise-consistent skew

In three linesNew TDA method for diffusion models using mirrored unlearning and noise-consistent skew. MUCS fine-tunes a second model with bounded mirrored gradient ascent and measures normalized skew using consistent noise samples. Outperforms existing approaches on three datasets.

Summary generated by Claude — human-verified