Reproduce “Restoring Diversity in Distilled Image and Video Generators by Probing Activations” 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

Furkan Horoz, Ozgur Kara, Kaan Oktay, Adil Kaan Akan, Atakan Kara, Anand Bhattad, James Matthew Rehg

ReDiverse restores variation in few-step image and video generators by targeting an intermediate block where student activations occupy fewer effective directions than the teacher’s. It uses teacher-derived directions through trained adapters or inference-time steering. Across four tested pairs, students retain or amplify total variance while retaining 37%–60% of teacher participation ratio. On SDXL, combined ReDiverse raises DINO diversity from 59.4 to 65.0, though quality measures are not uniformly unchanged.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.