PDMD: Projected Distribution Matching Distillation for Video Diffusion Models
What it is
This paper introduces Projected Distribution Matching Distillation (PDMD), a method to stabilize the training of video diffusion models by filtering out critic errors that cause sample degradation. It works by projecting out the component of the DMD update that is parallel to the student-critic endpoint residual, which is proven to be an unbiased estimate of the critic's error.
The problem it solves
Standard Distribution Matching Distillation (DMD) suffers from training instability where critic errors accumulate, leading to progressive oversaturation, unnatural textures, and degraded sample quality.
What you could use it for
- Accelerating video generation inference by reducing the number of function evaluations (NFE) to 4
- Improving the visual and audio quality of distilled video diffusion models like Wan2.1 and MiniMax-H3
- Stabilizing the training process of distribution matching distillation to prevent oversaturation and artifacts
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.