OPEN LINK ↗
// 1h agoRESEARCH PAPER
PDMD steadies few-step video diffusion
PDMD filters accumulated critic errors in distribution-matching distillation, improving four-step video and audio generation without extra losses, model passes, or training stages. The paper reports higher VBench and VideoGen-Eval scores than matched baselines, with open-source checkpoints and inference code ([paper](https://arxiv.org/abs/2609.35768), [repository](https://github.com/ZeamoxWang/pdmd)).
// ANALYSIS
PDMD targets a practical failure mode in fast video generation: distillation can become cheaper while quality progressively collapses. Its appeal is unusually strong because the fix is lightweight and directly compatible with existing DMD pipelines.
- –Projects away the update component aligned with estimated critic error, reducing oversaturation and unnatural textures.
- –Reports an 83.73 VBench score at four function evaluations with Wan2.1, 1.03 points above matched DMD.
- –Supports two- and four-step MiniMax-H3 video-audio checkpoints, including LoRA and full-weight variants.
- –Inference scripts are designed for a single 24GB or 80GB GPU, improving accessibility for experimentation.
- –The main caveat is that evidence remains benchmark- and model-specific; broader validation is needed before PDMD becomes a default distillation recipe.
// TAGS
pdmddistillationvideo-genaudio-geninferenceopen-sourceresearch
DISCOVERED
1h ago
2026-10-04
PUBLISHED
1h ago
2026-10-04
RELEVANCE
9/ 10
AUTHOR
AI Search