VA-TDM: A Unified View for Improving Few-Step Diffusion Models by Velocity Adversarial Distillation
The authors report state-of-the-art four-step image and video generation, including results on the 33-billion-parameter MiniMax-H3.
VA-TDM addresses a training mismatch: estimates of a few-step generator’s distribution can differ from the distribution actually used to update it. A learned critic compares student and teacher velocity fields, turning their disagreement into training feedback. The method supervises partial denoising along the student’s own generation trajectory. In human preference evaluations, the authors report that VA-TDM-H3 outperforms Fast-H3 and comes closer to the original many-step H3.
HF Daily Papers' note
VA-TDM addresses a training mismatch: estimates of a few-step generator’s distribution can differ from the distribution actually used to update it. A learned critic compares student and teacher velocity fields, turning their disagreement into training feedback. The method supervises partial denoising along the student’s own generation trajectory. In human preference evaluations, the authors report that VA-TDM-H3 outperforms Fast-H3 and comes closer to the original many-step H3.
HF Daily Papers' note
score 5