The authors start from the Distribution Matching Distillation (DMD) framework, which trains a student model to match the output distribution of a teacher diffusion model.
Predictive representation loss sharpens few-step video generation distillation
By injecting a self-supervised representation alignment term into distribution matching, the method reduces flickering and improves spatial detail across two video model architectures.
Chinese Tech
Junyi Chen · Haiyu Zhang · Zhoujie Fu · Tengfei Wang · Chunchao Guo
Shanghai Jiao Tong University · Hunyuan (Tencent) · Nanyang Technological University
Research Digest··3 min read
The authors identify that distribution matching distillation (DMD) for few-step video generation degrades because it only supervises output samples, lacking regularization on internal representations.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.