Streaming video generation has advanced through distribution matching distillation (DMD), a technique that aligns the joint distribution of video frames with a video teacher's approximation of the real video distribution. While this joint matching helps curb drift during autoregressive rollouts, the authors note that shortcomings persist in visual quality and semantic alignment.
To address these gaps, the researchers introduce DuoMatching, a distribution matching framework that approximates the real video distribution through a unified joint-marginal formulation. Beyond existing joint matching objectives, DuoMatching adds a marginal matching objective that supplies dedicated frame-level supervision from an image generator, transferring complementary visual and semantic priors from that image model.
Applying frame-level supervision to video generation raises a practical obstacle: a mismatch between the latent representations of the video student and the image teacher. The authors propose LatentBridge to resolve this mismatch. They also introduce Latent Variation Sampling, which spreads frame-level supervision across distinct temporal segments to reduce redundancy.
According to the paper, experiments show that DuoMatching improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluations report overall preference rates above 80% against all evaluated baselines. The authors state that the approach requires no extra inference compute, and that code has been released alongside a project page featuring video demos and side-by-side comparisons.