DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation
AuthorsJiahao Zhan, Yan Wang, Yongrui Ma, Qunliang Xing, Ruchang Yao, Runtao Liu, Shijie Zhao, Tianfan Xue
AffiliationsMMLab, CUHK · ByteDance Inc · Oct 2026 · HKUST · CPII under InnoHK
DuoMatching improves few-step video generation by combining video-level motion supervision with image-level visual and semantic guidance.
Key results
Aggregate score for two-step frame-wise generation.
Matched two-step frame-wise baseline score.
DuoMatching preference over Causal Forcing++.
Trainable parameters in the latent-mapping module.
Curated OpenVid videos used to train the module.
Temporal slices selected per video for marginal supervision.
What the paper found
In a ByteDance-linked study, DuoMatching improves few-step video generation by combining joint distribution matching from a video teacher with frame-level marginal matching from an image teacher. The joint objective helps preserve motion and temporal consistency; the marginal objective transfers visual detail and prompt-alignment priors. Two techniques make this combination practical: LatentBridge maps temporally compressed video latents into the image teacher’s latent space, while Latent Variation Sampling selects frames across changing temporal segments to reduce redundant supervision. With Qwen-Image as the default image teacher, the method scores 83.53 on VBench’s aggregate metric for two-step frame-wise generation, compared with 80.67 for Causal Forcing++. Human preference over Causal Forcing++ is 80.87% overall, and DuoMatching exceeds 80% against every tested baseline. LatentBridge has 10.75M parameters and is trained on 29,400 OpenVid videos; the default sampling budget is 4 latent slices per video. The approach also improves results from Wan2.1-T2V-1.3B without adding inference computation, while largely preserving motion quality.
Original abstract
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. To address these limitations, we propose DuoMatching, a distribution matching framework that approximates the real video distribution through a unified joint-marginal formulation. On top of existing joint matching formulations, the additional marginal matching objective provides dedicated frame-level supervision from an image generator, transferring complementary visual and semantic priors from it. To apply this frame-level supervision in video generation, we introduce LatentBridge to resolve the latent representation mismatch between the video student and the image teacher. Latent Variation Sampling further distributes such frame-level supervision across distinct temporal segments, reducing redundancy. Experiments demonstrate that DuoMatching improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluations show overall preference rates above 80% against all evaluated baselines. The project page is available at https://johnzhan2023.github.io/DuoMatching/.
Read the original paperMore in Diffusion Models
Browse all 64 papers →In-Distribution Forcing for Long Video Generation at Test Time
Jeongwoo Shin, Youngyoon Choi, Sangwoo Jo, Hyunmog Kim, Sungjoon Choi, Joonseok Lee, Jaewoong Choi, Jaemoo Choi
ID-Forcing stabilizes long video generation by ensuring that cached attention states remain consistent with the model’s training distribution.
S2PD: Serial-to-Parallel Diffusion for Physically and Logically Consistent Video Generation
Jeffrey Hu, Daniel Olmeda Reino, Ayush Tewari
S2PD first reasons through video events step by step, then refines them in parallel to make generated videos more physically and logically consistent without the full cost of autoregressive sampling.
ALoDLM: Adaptively Looped Diffusion Language Models
Liancheng Fang, Zhuowei Li, Youngeun Kim, Tianchen Zhao, Rajat Koner, Jiaye Wu, Linghan Xu, Xuanbai Chen, Xiang Xu, Zheng Zhang, Jakub Zablocki, Nishant Sankaran, Yifan Xing
ALoDLM makes diffusion language models smarter and faster by spending extra computation only on tokens that are difficult to predict.