Restoring Diversity in Distilled Image and Video Generators by Probing Activations
Distilled generators retain activation variance but concentrate it into too few directions, limiting how much outputs vary across random seeds.
The researchers locate this collapse in an intermediate block at the first denoising step. Their ReDiverse method restores missing variance directions through either a lightweight adapter trained against teacher activations or a training-free intervention using precomputed teacher principal directions. Across image and video distillation benchmarks, they report restored sample diversity while preserving generation quality, with support from a user study.
HF Daily Papers' note
The researchers locate this collapse in an intermediate block at the first denoising step. Their ReDiverse method restores missing variance directions through either a lightweight adapter trained against teacher activations or a training-free intervention using precomputed teacher principal directions. Across image and video distillation benchmarks, they report restored sample diversity while preserving generation quality, with support from a user study.
HF Daily Papers' note
score 5