DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation
Jiahao Zhan, Yan Wang, Yongrui Ma, Qunliang Xing, Ruchang Yao, Runtao Liu, Shijie Zhao, Tianfan Xue
CUHK · ByteDance Inc. · HKUST · CPII under InnoHK
Published October 2, 2026 · on the feed October 7, 2026 · ▲ 92 · arXiv 2610.03543
The 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/.
The Brief
2:10A short podcast summary of the paper. Written and voiced by AI.
Transcript
DuoMatching. Joint-Marginal Distribution Matching for Few-Step Video Generation. By Jiahao Zhan and seven others, from the Chinese University of Hong Kong, ByteDance, and two other institutions. Published on the second of October. Its code repository is worked out: about forty source files, with a readme and a licence.
Streaming video generation has gained a lot from distribution matching distillation, or DMD. The student is trained so that the joint distribution of its video frames matches a video teacher's approximation of the real video distribution. Matching the frames jointly helps against drift as the video is rolled out autoregressively. But it leaves limitations in two places: visual quality, and semantic alignment.
DuoMatching approximates the real video distribution through a unified formulation that is joint and marginal at once. The existing joint matching stays. On top of it sits an additional marginal matching objective, which supervises individual frames using an image generator. The image generator acts as a second teacher, and the frame-level supervision transfers its complementary visual and semantic priors into the video student.
Two pieces make that frame-level supervision usable for video. The first is LatentBridge. The video student and the image teacher do not share a latent representation, and LatentBridge resolves that mismatch so the image teacher's signal can reach the student at all. The second is Latent Variation Sampling, which spreads the frame-level supervision across distinct temporal segments of the video. Supervising different segments rather than the same ones reduces redundancy.
In experiments, DuoMatching improves visual quality, composition and semantic alignment, while largely preserving motion dynamics. Note the word largely: the frame-level gains are reported alongside motion that mostly, not entirely, holds its ground. In human evaluations, overall preference rates are above eighty percent against every baseline evaluated.
Each teacher here covers the other's blind side. A video teacher knows how frames go together, which keeps long rollouts from drifting. An image teacher knows what a single frame should look like and mean. DuoMatching lets the student learn from both, and the two engineering pieces exist so the image teacher can speak the student's latent language without repeating itself.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
The video flows. The frames, one by one, fall short.
You have a fast streaming video generator, distilled to a few steps by training it to match a video teacher over whole clips. Drift over long rollouts is under control.
The road ahead
- 1 · The problem ~6 min Training a fast video generator on whole clips keeps long videos steady, but it gives no direct reward for making any single frame better.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Keep whole-clip matching, and add an image generator as a second teacher that judges sampled single frames, read through a small learned bridge.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Pick slices from distinct segments, map them into image latents, let the image teacher judge them, and add that to whole-clip matching.
- 6 · The experiment ~6 min You’ll guess: In human evaluations, what overall preference rate did DuoMatching reach against the baselines it was compared with?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.