DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation
DuoMatching adds image-teacher supervision to video distillation so few-step generators keep stronger frame quality without giving up motion.
The paper says existing distribution matching distillation helps streaming video avoid rollout drift, but still falls short on visual quality and semantic alignment. DuoMatching combines joint video matching with a marginal frame-level objective drawn from an image generator. Its LatentBridge handles the mismatch between video-student and image-teacher latents, while Latent Variation Sampling spreads supervision across temporal segments. The authors report better visual quality, composition, and semantic alignment, with human preference rates above 80% against evaluated baselines. HF Daily Papers' note
The paper says existing distribution matching distillation helps streaming video avoid rollout drift, but still falls short on visual quality and semantic alignment. DuoMatching combines joint video matching with a marginal frame-level objective drawn from an image generator. Its LatentBridge handles the mismatch between video-student and image-teacher latents, while Latent Variation Sampling spreads supervision across temporal segments. The authors report better visual quality, composition, and semantic alignment, with human preference rates above 80% against evaluated baselines. HF Daily Papers' note
score 4