(DMD) reduces the number of denoising steps for video but can produce oversaturated, artifact-filled samples due to accumulating critic errors. Projected Distribution Matching Distillation (PDMD) addresses this by projecting out the component of the DMD update parallel to the student-critic endpoint residual, which is proven to be an unbiased estimate of the critic's endpoint error. This simple change stabilizes training and improves sample quality.
PDMD requires only a one-line code change to DMD with no extra losses or training stages. On Wan2.1 it achieves a VBench total score of 83.73 at 4 function evaluations, surpassing matched DMD by 1.03 points. On MiniMax-H3 joint video-audio generation it achieves a VideoGen-Eval visual total score of 83.17, 0.41 points above the strongest distilled baseline, and the best performance on all six audio metrics. Qualitative results and user studies confirm improved visual quality, motion, and audio.