Residual-direction validity for single-step video generation

Investigate whether removing the student–critic endpoint residual remains appropriate when a single function evaluation must cover the entire video generation process.

Background

The paper evaluates Projected Distribution Matching Distillation (PDMD) primarily for two- and four-step video generation. Although PDMD improves stability and quality in those settings, its single-step results remain unsatisfactory, with smeared samples and lower quality than the four-step outputs.

The authors specifically question whether the student–critic endpoint residual—the direction removed by PDMD—continues to identify a harmful component when one evaluation must represent the whole video-generation trajectory. Resolving this issue would clarify whether PDMD’s core projection mechanism can be extended reliably to one-step video distillation.