Lune

CVPR2026Top-tier venue

Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation

Yunhong Lu, Yanhong Zeng, Haobo Li, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jiapeng Zhu, Hengyuan Cao, Zhipeng Zhang, Xing Zhu, Yujun Shen, Min Zhang

Abstract

Efficient streaming video generation is critical for simulating interactive and dynamic worlds. Existing methods distill few-step video diffusion models with sliding window attention, using initial frames as sink tokens to maintain attention performance and reduce error accumulation. However, video frames become overly dependent on these static tokens, resulting in copied initial frames and diminished motion dynamics. To address this, we introduce Reward Forcing, a novel framework with two key designs. First, we propose EMA-Sink, which maintains fixed-size tokens initialized from initial frames and continuously updated by fusing evicted tokens via exponential moving average as they exit the sliding window. Without additional computation cost, EMA-Sink tokens capture both long-term context and recent dynamics, preventing initial frame copying while maintaining long-horizon consistency. Second, to better distill motion dynamics from teacher models, we propose a novel Rewarded Distribution Matching Distillation (Re-DMD). Vanilla distribution matching treats every training sample equally, limiting the model's ability to prioritize dynamic content. Instead, Re-DMD biases the model's output distribution toward high-reward regions by prioritizing samples with greater dynamics rated by a vision-language model. Re-DMD significantly enhances motion quality while preserving data fidelity. We include both quantitative and qualitative experiments to show that Reward Forcing achieves state-of-the-art performance on standard benchmarks while enabling high-quality streaming video generation at 23.1 FPS on a single H100 GPU.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

13 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Optimizing Few-Step Generation with Adaptive Matching Distillation cites Learning Transferable Visual Models From Natural Language Supervision
  • Mode Seeking meets Mean Seeking for Fast Long Video Generation cites Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention
  • Mode Seeking meets Mean Seeking for Fast Long Video Generation cites Scalable Diffusion Models with Transformers
  • Mode Seeking meets Mean Seeking for Fast Long Video Generation cites Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout
  • Improved Denoising Diffusion Probabilistic Models cites Denoising Diffusion Probabilistic Models
  • Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention cites Scalable Diffusion Models with Transformers
  • Scalable Diffusion Models with Transformers cites Improved Denoising Diffusion Probabilistic Models
  • Scalable Diffusion Models with Transformers cites Learning Transferable Visual Models From Natural Language Supervision
  • Scalable Diffusion Models with Transformers cites Denoising Diffusion Probabilistic Models
  • LoL: Longer than Longer, Scaling Video Generation to Hour cites Scalable Diffusion Models with Transformers
  • LoL: Longer than Longer, Scaling Video Generation to Hour cites Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout
  • Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout cites Scalable Diffusion Models with Transformers
  • Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout cites Denoising Diffusion Probabilistic Models
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers13

  1. Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-RolloutHidir Yesiltepe, Tuna Han Salih Meral, Adil Kaan Akan, Kaan Oktay et al.CVPR 2026 · 84 citations
  2. LoL: Longer than Longer, Scaling Video Generation to HourJustin Cui, Jie Wu, Ming Li, Tao Yang et al.CVPR 2026 · 30 citations
  3. Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse AttentionChengtao Lv, Yumeng Shi, Yushi Huang, Ruihao Gong et al.ICML 2026 · 10 citations
  4. Mode Seeking meets Mean Seeking for Fast Long Video GenerationShengqu Cai, Weili Nie, Chao Liu, Julius Berner et al.ICML 2026 · 9 citations
  5. Optimizing Few-Step Generation with Adaptive Matching DistillationLichen Bai, zikai ZHOU, Shitong Shao, Wenliang Zhong et al.ICML 2026 · 4 citations

Builds on54

  1. Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
  2. Denoising Diffusion Probabilistic ModelsJonathan Ho, Ajay Jain, Pieter AbbeelNeurIPS 2020 · 35,902 citations
  3. Direct Preference Optimization: Your Language Model is Secretly a Reward ModelRafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning et al.NeurIPS 2023 · 10,924 citations
  4. Scalable Diffusion Models with TransformersWilliam Peebles, Saining XieICCV 2023 · 5,568 citations
  5. Improved Denoising Diffusion Probabilistic ModelsAlexander Quinn Nichol, Prafulla DhariwalICML 2021 · 5,234 citations

Related papers

  1. Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video GenerationHongzhou Zhu, Min Zhao, Guande He, Hang Su et al.ICML 2026
  2. MotionStream: Real-Time Video Generation with Interactive Motion ControlsJoonghyuk Shin, Zhengqi Li, Richard Zhang, Jun-Yan Zhu et al.ICLR 2026 · 79 citations
  3. Rolling Forcing: Autoregressive Long Video Diffusion in Real TimeKunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan et al.ICLR 2026 · 215 citations
  4. Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative CompressionJung Yi, Wooseok Jang, Paul Cho, Jisu Nam et al.ICML 2026
  5. Transition Matching Distillation for Fast Video GenerationWeili Nie, Julius Berner, Nanye Ma, Chao Liu et al.CVPR 2026 · 24 citations
Dusk over the sea between two cliffs drawn in fine vertical lines