ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation
ViRDM post-trains few-step causal video generators without a teacher, scoring 84.87 on VBench.
ViRDM is a teacher- and critic-free post-training recipe for few-step causal video diffusion. It adapts representation distribution matching with stochastically truncated clean-exit supervision, a lightweight VAE decoder, staged vector-Jacobian products, and dynamics regularization. After 20 generator updates and 16 A100 GPU-hours, it scores 84.87 on official VBench, 0.36 above the previous best few-step causal baseline. The authors also report exploratory results for lower sampling budgets and one-, two-, and four-step bidirectional generation.
- Drops the pretrained teacher and online critic required by distribution matching distillation.
- Uses truncated supervision, a light VAE decoder, and staged vector-Jacobian products.
- Scores 84.87 on VBench after 20 updates and 16 A100 GPU-hours.
- Adds dynamics regularization because image-style matching underconstrains video motion.
From the original article
1 min read at huggingface.co
Few-step autoregressive (AR) video diffusion enables low-latency streaming generation, but existing post-training methods predominantly rely on Distribution Matching Distillation (DMD), requiring both a large pretrained teacher and an online critic to estimate distributional discrepancies through diffusion scores.
By Zichong Meng, Chongjian Ge, Chun-Hao P. Huang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.