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

cs.CVOct 2, 2026

DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation

Authors: Jiahao Zhan, Yan Wang, Yongrui Ma, Qunliang Xing, Ruchang Yao, Runtao Liu, Shijie Zhao, Tianfan Xue

Organizations: MMLab, CUHK · ByteDance Inc. · HKUST · CPII under InnoHK

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/.

Figures & tables

Figure 1: Motivation and results of DuoMatching. Videos generated by autoregressive models often have degraded quality for later rollout frames (see Q-Align [ 38 ] curves). Joint DMD mitigates this drift, yet the student still retains a visual quality gap relative to the video teacher. The example reveals a missing feather specified in the prompt. To solve this problem, we reformulate distribution matching from a coupled joint-marginal perspective to better approximate the real video distribution, as illustrated by improved visual quality and the recovered feather, while maintaining temporal coherence.

Figure 2: Spatial structure of score differences in joint and marginal distribution matching. Marginal matching exhibits more spatially coherent responses, whereas joint matching produces more scattered responses in this example. Each heatmap is independently normalized to highlight spatial structure rather than absolute magnitude.

Figure 3: Overview of DuoMatching. (a) DuoMatching complements joint distribution matching with an image-teacher reference distribution $P_{i}$ . Each group schematically represents a video, with nodes depicting frame-level distributions. (b) LatentBridge is trained to map a video latent $z^{l}$ , conditioned on the preceding slice $z^{l-1}$ and a local frame index $i$ , to the corresponding image latent using an $\ell_{1}$ reconstruction loss. (c) Latent Variation Sampling uses differences between adjacent latent slices to partition the sequence into segments, within which slices are randomly sampled.

Figure 4: Qualitative comparison between DuoMatching and Causal Forcing++ across four aspects: fine-grained visual quality, semantic alignment, composition, and dynamics. DuoMatching improves visual details, prompt adherence, and composition while preserving video dynamics.

Bidirectional DMD methods
Method	Mode	NFE	Semantic $\uparrow$	Aesthetic $\uparrow$	Imaging $\uparrow$	Dynamic $\uparrow$	Smoothness $\uparrow$	Total $\uparrow$
CausVid	Full	4	69.23	64.15	73.00	36.67	98.84	80.63
Ours	Full	4	70.89	64.55	73.77	39.17	98.82	81.45
Causal autoregressive methods
Frame-wise generation (1 step)
Causal Forcing++	Frame	1	57.17	67.12	68.83	70.56	98.49	79.82

Table 1: Quantitative comparison on VBench with Self Forcing [ 11 ] , One-Forcing [ 7 ] , Reward Forcing [ 21 ] and Causal Forcing++ [ 47 ] . “Full” denotes bidirectional full-video generation, while “Frame” and “Chunk” denote the causal autoregressive generation unit. NFE denotes the number of function evaluations per generation unit. Results should be compared within matched generation modes and NFE budgets. Bold indicates the best result within each matched group; ours are shaded in light gray.

Comparison	Visual Quality $\uparrow$	Semantic Alignment $\uparrow$	Temporal & Motion $\uparrow$	Overall $\uparrow$
Ours vs. CausVid	92.17%	89.13%	51.30%	91.30%
Ours vs. Causal Forcing++	81.30%	79.57%	49.57%	80.87%
Ours vs. One-Forcing	93.48%	91.74%	90.87%	94.35%
Ours vs. Reward Forcing	83.91%	82.61%	60.87%	83.91%

Table 2: Human preference study under the same Wan2.1 video teacher. Each entry reports the percentage of pairwise judgments favoring our method over the corresponding baseline.

	Teacher	Video Student
Image Teacher	HPSv3 $\uparrow$	Semantic $\uparrow$	Aesthetic $\uparrow$	Imaging $\uparrow$	Dynamic $\uparrow$	Smoothness $\uparrow$	Total $\uparrow$
None	–	68.42	62.28	69.82	94.44	98.20	80.67
Wan2.1-14B	5.87	68.77	64.16	69.90	94.72	99.05	81.43
SDXL	6.57	68.50	64.78	70.01	93.89	98.12	81.47
FLUX.2-4B	8.86	69.06	65.29	71.58	94.17	99.23	82.64
Qwen-Image	9.05	72.84	66.97	73.48	93.61	98.39	83.53

Table 3: Ablation of image teachers, comparing Wan2.1-14B [ 32 ] , SDXL [ 25 ] , FLUX.2-4B [ 14 ] , and Qwen-Image [ 37 ] against a baseline without an image teacher. Teacher HPSv3 [ 22 ] measures human preference alignment of images generated by each teacher, while the remaining metrics evaluate the resulting video student.

Configuration		Evaluation Metrics				Peak Memory
Marginal DMD	Latent Mapping	Semantic $\uparrow$	Imaging $\uparrow$	Dynamic $\uparrow$	Total $\uparrow$	(GB) $\downarrow$
$\times$	—	68.42	69.82	94.44	80.67	56.49
$\checkmark$	Direct	71.74	71.66	76.94	81.78	59.40
$\checkmark$	Decode-Encode	—	—	—	—	OOM
$\checkmark$	LatentBridge	72.84	73.48	93.61	83.53	59.41

Table 4: Ablation of marginal DMD and latent mapping strategies. Direct applies marginal DMD directly to temporally compressed video latent slices. Decode-Encode decodes video latents into RGB frames and re-encodes sampled frames with the image VAE, retaining gradients through both operations. Peak memory denotes the maximum GPU memory allocated to tensors during a complete training iteration. OOM indicates an out-of-memory failure.

Sampling strategy	$K=2$	$K=4$	$K=8$
	Dynamic $\uparrow$ / Total $\uparrow$	Dynamic $\uparrow$ / Total $\uparrow$	Dynamic $\uparrow$ / Total $\uparrow$
Uniform random	93.06 / 80.99	91.39 / 82.61	89.44 / 81.04
Temporally stratified	– / –	91.67 / 82.68	– / –
LVS (Ours)	94.17 / 81.15	93.61 / 83.53	92.22 / 82.55

Table 5: Ablation of the number of sampled latent slices

K

and sampling strategies. Temporally stratified sampling randomly selects one latent slice from each of

K

segments of equal length.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	$\omega=0.1$	$\omega=0.4$	$\omega=0.8$
Semantic $\uparrow$	69.60	72.84	71.49
Imaging $\uparrow$	70.79	73.48	71.74
Dynamic $\uparrow$	93.61	93.61	87.78
Total $\uparrow$	81.03	83.53	81.57

Appendix

Table 6: Ablation of the marginal DMD loss weight

\omega

Figure 5: Additional qualitative comparisons of DuoMatching with Reward Forcing, One-Forcing, and CausVid. DuoMatching improves visual details, semantic alignment, and composition while preserving video dynamics.

Figure 6: Qualitative evaluation of LatentBridge across image latent spaces. Top: target latents obtained by encoding the corresponding RGB frames with each image VAE. Bottom: reconstructions from latents predicted by LatentBridge from video latents of Wan2.1. Both rows use the corresponding image VAE decoder.

Figure 7: Latent reconstruction error of LatentBridge at different stages of generator training. Given a generated video latent slice

z^{l}

, with

z^{l-1}

and the local frame index

i

as conditioning, LatentBridge predicts the corresponding image latent. We measure its

\ell_{1}

error against the target obtained through Decode-Encode over 100 samples.

Single-Step Video Generation

See all 11 papers

Joint Alignment and Distillation for Video Generation via Sample-Guided Distribution Matching
Sep 3, 2026Jiuzhou Lin, Junlong Wu, Fei Zuo +11Video Generation Diffusion Model Alignment
Tsinghua University · Kuaishou Technology · Harbin Institute of Technology (ShenZhen) +1
VDOT++: Unified Few-Step Video Generation via Unbalanced Optimal Transport Distillation
Oct 2, 2026Yutong Wang, Xingtong Ge, Enhuai Liu +6Diffusion Model Distillation Few-Step Diffusion Sampling
University of Sydney · Hong Kong University of Science and Technology · Shanghai AI Laboratory +1
Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation
Apr 3, 2026Xingtong Ge, Yi Zhang, Yushi Huang +6Diffusion Model Distillation Video Generation
Hong Kong University of Science and Technology · Vivix Group Limited

Sep 3, 2026J/K move · Enter open · S save

Joint Alignment and Distillation for Video Generation via Sample-Guided Distribution Matching

Jiuzhou Lin, Junlong Wu, Fei Zuo, Huan Ouyang+10

Tsinghua University · Kuaishou Technology · Harbin Institute of Technology (ShenZhen)+1

Video Generation Diffusion Model Alignment Distribution Matching Distillation Video Diffusion Models Diffusion Model Distillation

Open paper AskPDF arXiv

Aligning video generative models to human preferences heavily relies on Reinforcement Learning (RL), which suffers from extensive computational overhead. Existing workflows typically treat RL and distillation as disconnected stages: applying RL before distillation incurs prohibitive computational costs, whereas applying RL after distillation frequently leads to model collapse. To overcome these limitations, we propose a unified, single-stage optimization framework grounded in Distribution Matching (DM). In the standard DM framework, distillation updates the model via a gradient direction that minimizes the gap between the real and fake models, guiding generations toward clarity and high fidelity. Building upon this, we introduce DM-Align, which derives a complementary gradient direction to guide the model toward human-preferred samples. Inspired by DPO and GRPO, our method leverages the distributional gap -- formulated from either preference pairs or intra-group exploration -- to directly construct this preference-guided gradient. By synergizing these two gradient directions, our approach eliminates the need for multi-step reward evaluation and complex ODE-SDE conversions inherent in traditional RL. Comprehensive experiments across multiple foundational video models demonstrate that this sample-guided framework robustly enhances both distillation quality and preference alignment, consistently outperforming both standalone variants and sequential two-stage pipelines.

Abstract

Single-Step Video Generation

Explore similar work

Joint Alignment and Distillation for Video Generation via Sample-Guided Distribution Matching

VDOT++: Unified Few-Step Video Generation via Unbalanced Optimal Transport Distillation

Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation