Few-Step Diffusion Distillation
Methods for distilling diffusion models into a small number of denoising steps while maintaining generation fidelity and stability.
7 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · ByteDance Inc., University of California, San Diego
Projecting critic errors stabilizes four-step video diffusion distillation
Identifies critic estimation errors as the cause of oversaturation in distribution matching distillation and stabilizes four-step video diffusion generation.
4 further papers
Oct 2026 · Inria, ENS de Lyon
Spectral hybrid parameterizations robustly accelerate flow matching optimization
Analyzes why different flow matching prediction targets (x1, v, x0) yield different performance and proposes a spectral hybrid parameterization that accelerates training.
Oct 2026 · Laboratoire de Physique de l’École normale supérieure, ENS, Université PSL, CNRS, Sorbonne Université, Université Paris Cité
Weighting schedules in score-based models determine which data features are learned and at what rate
Connects training-dynamics weighting schedules to which data features are learned, offering a principled guide for distilling score-based models into fewer steps.
3 of 7 papers shown