DuoMatching reports 80%+ human preference over video baselines

A distillation method called DuoMatching reports human preference rates "above 80%" against every baseline it was evaluated on, in a paper posted to arXiv on October 2 by a group led by Jiahao Zhan.

The approach builds on distribution matching distillation (DMD), which has driven recent progress in streaming video generation by letting a student model match the joint distribution of frames predicted by a video teacher. The authors argue joint matching alone leaves gaps in visual quality and semantic alignment, so they bolt on a second signal: "the additional marginal matching objective provides dedicated frame-level supervision from an image generator," the abstract says, transferring per-frame priors into a model that otherwise only sees video-level targets. Two components handle the practical plumbing. LatentBridge reconciles the latent spaces of the video student and the image teacher, and Latent Variation Sampling spreads the per-frame supervision across temporal segments so it does not pile up on the same regions.

The abstract claims the method "improves visual quality, composition, and semantic alignment while largely preserving motion dynamics." That "largely" is the paper's own hedge. Motion is the axis where image-level supervision could plausibly cost something, and the abstract publishes no per-axis numbers and does not name the baselines it beats.