DuoMatching improves few-step video generation by combining joint and marginal distribution matching

Few-step video generation aims to produce videos with high visual quality and temporal coherence using a small number of inference steps, which is critical for real-time applications like interactive simulations and digital entertainment. However, existing video generation models often require many sequential denoising steps, leading to high computational cost and latency.

Distribution matching distillation (DMD) has been used to distill complex multi-step video generation models into few-step generators by matching the joint distribution of video frames between the student and a pretrained video teacher. This joint matching helps reduce drift during autoregressive rollouts but still leaves limitations in visual quality and semantic alignment, especially at the frame level.

Existing DMD methods focus on matching the joint distribution of all frames, which constrains cross-frame dependencies but does not provide dedicated supervision for individual frames. This can cause frame-level visual deficiencies, such as missing fine details or semantic inconsistencies, because improving one frame might conflict with others under joint constraints.