DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation
DuoMatching adds joint-marginal matching so few-step video generators improve visual quality and semantic alignment.
DuoMatching is a few-step video distillation method that matches both the joint video distribution and per-frame marginals from an image generator. LatentBridge resolves latent mismatch between the video student and image teacher, and Latent Variation Sampling spreads frame-level supervision across temporal segments. The authors report better visual quality, composition, and semantic alignment while largely keeping motion, with human preference above 80% against baselines.
- Adds marginal frame matching on top of joint video distribution matching.
- LatentBridge aligns video-student and image-teacher latent spaces.
- Latent Variation Sampling spreads frame supervision across temporal segments.
- Human preference exceeds 80% versus evaluated baselines.
From the original article
1 min read at huggingface.co
Streaming video generation has benefited from distribution matching distillation (DMD), which matches the joint distribution of video frames to a video teacher's approximation of the real video distribution. Although this joint matching mitigates drift during autoregressive rollouts, limitations remain in visual quality and semantic alignment.
By Jiahao Zhan, Yan Wang, Yongrui Ma · 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.