Diffusion Distillation
Methods for distilling diffusion and flow-matching models to reduce inference cost without quality loss.
10 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · The Hong Kong Polytechnic University, OPPO Research Institute
Splitting diffusion distillation into two phase-specific half-sized experts beats a full-size student.
Phase-wise Velocity Distillation splits generation into two phases with half-sized experts, outperforming a full-size student.
2 further papers
Oct 2026 · EPFL, NVIDIA
Polynomial student paths improve four-step video distillation quality
Introduces parametric trajectory distillation for four-step video generation, improving on prior one-step methods.
Oct 2026 · Amazon AGI, Purdue University
Consistency error enables direct unlearning in few-step distilled diffusion models
Introduces CePU, a preference-driven unlearning method that uses consistency error for few-step distilled diffusion models.
4 further papers
Oct 2026 · Together AI, University of Illinois Urbana-Champaign
On-policy residual training helps diffusion language models decode faster
Adds on-policy residual training that reuses and refines earlier denoising distributions to speed up diffusion language model decoding.
4 of 10 papers shown