PDMD: Projected Distribution Matching Distillation for Video Diffusion Models
PDMD is presented as a one-line change to DMD that filters critic errors and stabilizes 4-step video generation.
The paper says DMD can degrade during training because critic errors accumulate across student updates, leading to oversaturation and artifacts. PDMD removes the update component aligned with the student-critic endpoint residual, which the authors argue estimates critic endpoint error. In tests on Wan2.1, it reports a VBench score of 83.73 at 4 NFE, 1.03 points above matched DMD. On MiniMax-H3 video-audio generation, it also leads the compared 4-NFE models across all six audio metrics. ArXiv · AI/CL/LG's note
The paper says DMD can degrade during training because critic errors accumulate across student updates, leading to oversaturation and artifacts. PDMD removes the update component aligned with the student-critic endpoint residual, which the authors argue estimates critic endpoint error. In tests on Wan2.1, it reports a VBench score of 83.73 at 4 NFE, 1.03 points above matched DMD. On MiniMax-H3 video-audio generation, it also leads the compared 4-NFE models across all six audio metrics. ArXiv · AI/CL/LG's note
score 5