Paper 2605.03849
Stream-R1: Reliability-Perplexity Aware Reward Distillation for Streaming Video Generation
- Published
- May 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- 54 stars
01 In brief
Summary
Stream-R1 is a framework for distilling autoregressive streaming video diffusion models, addressing limitations in existing distribution matching distillation (DMD) methods that treat all rollouts, frames, and pixels equally.
It introduces two concepts: Inter-Reliability (varying reliability of supervision across rollouts) and Intra-Perplexity (varying potential for improvement across spatiotemporal regions).
Stream-R1 uses a single pretrained video reward model to compute an Inter-Reliability weight (exponential of reward score) and an Intra-Perplexity weight map (from reward gradient saliency, decomposed into spatial and temporal components).
An adaptive balancing mechanism prevents any quality dimension (visual, motion, text alignment) from dominating.
Experiments on VBench show Stream-R1 achieves the highest Total score (84.40) among compared methods, surpassing its teacher Wan2.1 (84.26) and baseline Reward Forcing (84.13), with improvements in Quality and Semantic scores.
On long videos (up to 180s), it outperforms Reward Forcing on all metrics, with the gap widening at longer durations.
Human preference evaluation shows Stream-R1 preferred on all five dimensions, especially Dynamic Reasonableness (63% win rate) and Visual Quality (60%).
Ablations confirm each component contributes, with temporal decomposition providing the largest gain.
The method adds no inference cost and requires no architectural changes to the student model.
Training used Wan2.1-T2V-1.3B as student and 14B as teacher, with 8 A100 GPUs for about 56 hours.
The paper is dated May 6, 2026, and the project page is https://stream-r1.github.io.
Correspondence: Mengqi Huang at huangmq@ustc.edu.cn.
The work is by Bin Wu, Mengqi Huang, Shaojin Wu, Weinan Jia, Yuxin Wang, Zhendong Mao, and Yongdong Zhang from USTC, FrameX.AI, and…
02 From the paper
Abstract
Distillation-based acceleration has become foundational for making autoregressive streaming video diffusion models practical, with distribution matching distillation (DMD) as the de facto choice. Existing methods, however, train the student to match the teacher's output indiscriminately, treating every rollout, frame, and pixel as equally reliable supervision. We argue that this caps distilled quality, since it overlooks two complementary axes of variance in DMD supervision: Inter-Reliability across student rollouts whose supervision varies in reliability, and Intra-Perplexity across spatial regions and temporal frames that contribute unequally to where quality can still be improved. The objective thus conflates two questions under a uniform weight: whether to learn from each rollout, and where to concentrate optimization within it. To address this, we propose Stream-R1, a Reliability-Perplexity Aware Reward Distillation framework that adaptively reweights the distillation objective at both rollout and spatiotemporal-element levels through a single shared reward-guided mechanism. At the Inter-Reliability level, Stream-R1 rescales each rollout's loss by an exponential of a pretrained video reward score, so that rollouts with reliable supervision dominate optimization. At the Intra-Perplexity level, it back-propagates the same reward model to extract per-pixel gradient saliency, which is factored into spatial and temporal weights that concentrate optimization pressure on regions and frames where refinement yields the largest expected gain. An adaptive balancing mechanism prevents any single quality axis from dominating across visual quality, motion quality, and text alignment. Stream-R1 attains consistent improvements on all three dimensions over distillation baselines on standard streaming video generation benchmarks, without architectural modification or additional inference cost.