Mode Seeking meets Mean Seeking for Fast Long Video Generation

Shengqu Cai; Weili Nie; Chao Liu; Julius Berner; Lvmin Zhang; Nanye Ma; Hansheng Chen; Maneesh Agrawala

cs.CVFeb 27, 2026

Mode Seeking meets Mean Seeking for Fast Long Video Generation

Authors: Shengqu Cai, Weili Nie, Chao Liu, Julius Berner, Lvmin Zhang, Nanye Ma, Hansheng Chen, Maneesh Agrawala,

Organizations: Stanford University, California, USA · NVIDIA Research, California, USA · NYU Courant, New York, USA.

Abstract

Scaling video generation from seconds to minutes faces a critical bottleneck: while short-video data is abundant and high-fidelity, coherent long-form data is scarce and limited to narrow domains. To address this, we propose a training paradigm where Mode Seeking meets Mean Seeking, decoupling local fidelity from long-term coherence based on a unified representation via a Decoupled Diffusion Transformer. Our approach utilizes a global Flow Matching head trained via supervised learning on long videos to capture narrative structure, while simultaneously employing a local Distribution Matching head that aligns sliding windows to a frozen short-video teacher via a mode-seeking reverse-KL divergence. This strategy enables the synthesis of minute-scale videos that learns long-range coherence and motions from limited long videos via supervised flow matching, while inheriting local realism by aligning every sliding-window segment of the student to a frozen short-video teacher, resulting in a few-step fast long video generator. Evaluations show that our method effectively closes the fidelity-horizon gap by jointly improving local sharpness, motion and long-range consistency. Project website: https://primecai.github.io/mmm/.

Figures & tables

Figure 1: Video length is not analogous to image resolution. Top: For images, moving from low to high resolution is largely an interpolation of the same underlying local patch distribution. Bottom: For videos, moving from short clips to long sequences is temporal extrapolation. The model must introduce new events and causal structure beyond the short-clip horizon, which is fundamentally harder than multi-resolution image training.

Figure 2: Overview: mode seeking meets mean seeking. A shared long-context condition encoder $E_{\phi}$ maps a noisy long-video latent $x_{t}^{\text{long}}$ (with timestep $t$ and conditioning $c$ ) to a unified representation $h_{t}$ . Two lightweight decoder heads read out velocities from $h_{t}$ : the long context Flow Matching head $D^{\text{FM}}_{\theta}$ is trained with supervised flow matching on real long videos (mean-seeking), while the segment-wise Distribution Matching head $D^{\text{DM}}_{\psi}$ is trained via on-policy sliding-window reverse-KL alignment to an expert short-video teacher using DMD ( Yin et al., 2024b ; Yin et al., 2024a ) /VSD ( Wang et al., 2023 ) -style gradients (mode-seeking). Both objectives update the shared encoder, but each head receives only its corresponding signal.

Method	NFE	Subject $\uparrow$	Background $\uparrow$	Motion $\uparrow$	Dynamic $\uparrow$	Aesthetic $\uparrow$	Image $\uparrow$	VLM $\uparrow$
Method	NFE	Consistency	Consistency	Smoothness	Degree	Quality	Quality	Consistency
Long-context SFT	50	0.9685	0.9533	0.9866	0.9375	0.4973	0.6303	77.28
Mixed-length SFT	50	0.9667	0.9541	0.9874	0.8906	0.5467	0.6683	74.63
CausVid ( Yin et al., 2025 )	4	0.9736	0.9614	0.9789	0.8594	0.6044	0.6305	39.30
Self Forcing ( Huang et al., 2025b )	4	0.9489	0.9451	0.9805	0.9063	0.5556	0.6278	37.60
InfinityRoPE ( Yesiltepe et al., 2026 )	4	0.9689	0.9573	0.9812	0.7188	0.5342	0.6871	68.61

Table 1: Quantitative comparisons. The first , second , and third values are highlighted. NFE is reported for diffusion/flow samplers in our setup; LongLive ( Yang et al., 2026 ) is a frame-level AR real-time system, so its NFE is not directly comparable. We note that rollout-based methods can get over-saturated (CausVid ( Yin et al., 2025 ) ) or static/conservative (InfinityRoPE ( Yesiltepe et al., 2026 ) , LongLive ( Yang et al., 2026 ) ), which in turn significantly boosts some consistency metrics. Even so, our method still stands out as the best overall performing model.

Figure 3: Qualitative results. Our method generalizes well to various scenarios, producing long videos that maintain local fidelity and global coherence. All results are obtained using the Wan ( Team Wan et al., 2025 ) 1.3B model as both the student and the teacher, demonstrating how our decoupled training effectively extends short-video capabilities to long-horizon generation. We refer to our supplemental website for the videos and results from the Wan ( Team Wan et al., 2025 ) 14B model.

Figure 4: Qualitative comparison. “LongSFT” and “MixSFT” refer to long-context supervised finetuning (SFT) and mixed-length SFT, respectively. SFT-only methods (LongSFT, MixSFT) achieve decent long context and narrative, but appear to be blurry. Teacher-only methods (CausVid ( Yin et al., 2025 ) , Self-Forcing ( Huang et al., 2025b ) ) suffer from quality degradation over long range. While InfinityRoPE ( Yesiltepe et al., 2026 ) extends the temporal horizon, it does not have the ability to process long context and tends to generate still contents. Our method stands out as the best-performing model overall in terms of quality, motion, and long-horizon consistency. We refer to our supplementary website for video results, including extended comparisons with LongLive ( Yang et al., 2026 ) .

Method	Consistency $\uparrow$	Motion $\uparrow$	Quality $\uparrow$
No DDT dual heads	0.9427	0.9449	0.5298
No Sliding-window DMD	0.9604	0.9621	0.6075
No SFT	0.9579	0.9690	0.5862
Full Model	0.9615	0.9685	0.6359

Table 2: Ablation study. The first , second , and third values are highlighted. Our full model achieves the best overall performance, highlighting the necessity of all components.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Subject	Background	Motion	Dynamic	Aesthetic	Image	VLM
Method	Consistency $\uparrow$	Consistency $\uparrow$	Smoothness $\uparrow$	Degree $\uparrow$	Quality $\uparrow$	Quality $\uparrow$	Consistency $\uparrow$
CausVid ( Yin et al., 2025 )	0.9702	0.9589	0.9771	0.5078	0.5596	0.6264	41.82
Self-Forcing ( Huang et al., 2025b )	0.9516	0.9467	0.9784	0.4609	0.5524	0.6291	40.37
InfinityRoPE ( Yesiltepe et al., 2026 )	0.9683	0.9568	0.9810	0.7891	0.5415	0.6826	66.08
LongLive ( Yang et al., 2026 )	0.9679	0.9572	0.9827	0.8203	0.5704	0.6908	62.74
Ours	0.9676	0.9569	0.9839	0.9297	0.5861	0.7015	73.08

Appendix

Table 3: Additional 60-second quantitative comparison. The first , second , and third values are highlighted within this 60-second stress test. Our method achieves the strongest motion smoothness, dynamic degree, aesthetic quality, image quality, and VLM consistency, while remaining competitive on subject/background consistency.

Horizon	Method	Overall Quality $\uparrow$	Global Consistency $\uparrow$	Static Artifact $\downarrow$
30 sec	LongSFT	2.85	3.09	2.11
	MixedSFT	2.66	3.42	1.64
	CausVid ( Yin et al., 2025 )	1.98	1.92	2.98
	Self-Forcing ( Huang et al., 2025b )	2.71	1.75	2.82
	InfinityRoPE ( Yesiltepe et al., 2026 )	3.23	2.99	4.12
	LongLive ( Yang et al., 2026 )	3.12	3.27	4.03

Appendix

Table 4: User study. The first , second , and third values are highlighted, with the ordering reversed for static artifacts because lower is better. Our method receives the highest overall quality and global consistency ratings for both 30-second and 60-second videos, and substantially reduces static artifacts relative to rollout-based baselines.

Video Diffusion Models

Video diffusion models generate or edit temporally coherent video by iteratively denoising latent sequences conditioned on text, images, motion, or other signals.

See all 735 papers

Towards Error-Free Long Video Generation
Jun 21, 2026Shuning Chang, Weihua Chen, Jiasheng Tang +8Video Diffusion Models Long Video Generation
Zhejiang University · Alibaba Group
Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model
Aug 6, 2026Haoning Yang, Xinyuan Chen, Yaohui Wang +1Diffusion Model Guidance Video Diffusion Models
Shanghai Jiao Tong University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China
FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching
May 20, 2026Jangho Park, Geon Yeong Park, Gihyun Kwon +1Video Diffusion Models Long Video Generation
1KAIST · 2Amazon

Jun 21, 2026J/K move · Enter open · S save

Towards Error-Free Long Video Generation

Shuning Chang, Weihua Chen, Jiasheng Tang, Hao Xu+7

Zhejiang University · Alibaba Group

Video Diffusion Models Long Video Generation Autoregressive Video Generation Temporal Consistency in Video Generation

Open paper AskPDF arXiv

Recent advances in video generation have made minute-level synthesis possible; however, generating long videos remains challenging due to error accumulation, attribute drift, and the limited availability of long video data. In this paper, we introduce an infinite-length video generation framework that focusing on addressing these issues and produces high-quality, dynamic, and identity-consistent single-shot long videos. We first finetune a diffusion model as a video extension model on large-scale short video data to autoregressively generate temporally coherent clips. Inspired by the success of large language models (LLMs), we adopt causal attention computation between clips to further finetune this model on long video data. In this way, the tokens in one clip (short video) are computed by bidirectional attention while tokens among clips are computed by unidirectional attention. This design leverages the strengths of modern diffusion models while preserving long-term context information, effectively mitigating error accumulation and attribute drift. To achieve memory efficiency during inference, we adopt a key-value (KV) caching mechanism to maintain a constant KV memory. Furthermore, we introduce truncation-rectified flow (T-RFlow) technique to further suppress error accumulation. Experimental results demonstrate the effectiveness of our method. Our framework establishes a new benchmark for realistic and coherent minute-level video synthesis.

Abstract

Video Diffusion Models

Explore similar work

Towards Error-Free Long Video Generation

Diff-VF: Training-free High-quality Long Video Generation via Diffusion Model

FlowLong: Inference-time Long Video Generation via Manifold-constrained Tweedie Matching