SGMD: Score Gradient Matching Distillation for Few-Step Video Diffusion Distillation

Researchers introduced Score Gradient Matching Distillation (SGMD) for few-step video diffusion.

SGMD directly matches a fake score to the teacher with stop-gradient Fisher matching and dual potentials for better outer-loop correction and inner-loop tracking.

Compared with DMD2, it trains about 3x faster and produces stronger motion dynamics in 4-step distilled models.

The method preserves temporal consistency and maintains visual quality and text alignment.

This could make practical video generation systems faster and more effective.