Hover any clip to pause · click to play fullscreen
Abstract
Combined with candidate selection, Stream-T1 actively optimizes the generation trajectory by dynamically refining both the latent noise and the context memory.
While Test-Time Scaling (TTS) offers a promising direction to enhance video generation without the surging costs of training, current test-time video generation methods based on diffusion models suffer from exorbitant candidate exploration costs and lack temporal guidance. To address these structural bottlenecks, we propose shifting the focus to streaming video generation. We identify that its chunk-level synthesis and few denoising steps are intrinsically suited for TTS, significantly lowering computational overhead while enabling fine-grained temporal control. Driven by this insight, we introduced Stream-T1, a pioneering comprehensive TTS framework exclusively tailored for streaming video generation. Evaluated on both 5s and 30s comprehensive video benchmarks, Stream-T1 demonstrates profound superiority, significantly improving temporal consistency, motion smoothness, and frame-level visual quality.
actively refines the initial latent noise of the generating chunk using historically proven, high-quality previous chunk noise, effectively establishes temporal dependency and utilizing the historical Gaussian prior to guide the current generation;
comprehensively evaluates generated candidates to strike an optimal balance between local spatial aesthetics and global temporal coherence by integrating immediate short-term assessments with sliding-window-based long-term evaluations;
dynamically routes the context evicted from KV-cache into distinct updating pathways guided by the reward feedback, ensuring that previously generated visual information effectively anchors and guides the subsequent video stream.
Method
For each chunk, Stream-T1 performs optimization in the following order: pre-generation, post-generation, and post-pruning.
Video Generation
Each row plays clips that were independently rolled out at a different duration — 5 seconds, 30 seconds, Hover a row to pause its scroll and zoom the clip under your cursor; click any clip to play it fullscreen.
Method Comparison
Same prompt, same seed. Stream-T1 significantly elevates the temporal consistency and frame-level visual fidelity of the generated videos.
Full search visualization
Quantitative Results
To comprehensively evaluate the effectiveness of our proposed Stream-T1, we compare against three representative open-source models on both 5s and 30s video generation. Overall, extensive experimental analyses demonstrate that Stream-T1 significantly and comprehensively elevates the temporal consistency, motion coherence, and frame-level visual fidelity of the generated videos.
Stream-T1 records the best results on six quality metrics(Subject Consistency, Background Consistency, Motion Smoothness, Aesthetic Quality, MQ and TA) and ranks the second on Imaging Quality and VQ.
| Method | VBench↑ | VideoAlign↑ | ||||||
|---|---|---|---|---|---|---|---|---|
| Subject Consistency |
Background Consistency |
Motion Smoothness |
Imaging Quality |
Aesthetic Quality |
VQ | MQ | TA | |
| CausVid | 96.33 | 95.56 | 98.66 | 69.69 | 62.90 | 0.433 | 0.550 | 1.02 |
| Self-forcing | 95.26 | 95.67 | 98.67 | 71.61 | 63.97 | 0.099 | 0.088 | 1.193 |
| LongLive | 97.00 | 96.78 | 99.12 | 71.28 | 65.28 | 0.285 | 0.350 | 1.193 |
| Stream-T1 (on LongLive) | 97.25 | 97.05 | 99.15 | 71.42 | 65.98 | 0.426 | 0.629 | 1.305 |
| Relative gain | Δ0.26% | Δ0.28% | Δ0.03% | Δ0.20% | Δ1.07% | Δ49.47% | Δ79.71% | Δ9.39% |
Stream-T1 outperforms state-of-the-art baselines across almost all metrics. Stream-T1 achieves the highest scores in Subject Consistency, Background Consistency, Motion Smoothness, Imaging Quality, and Aesthetic Quality. This superiority is further corroborated by the human-aligned VideoAlign, where our model obtains the best VQ and TA. And Stream-T1 gets second score in MQ.
| Method | VBench Long↑ | VideoAlign↑ | ||||||
|---|---|---|---|---|---|---|---|---|
| Subject Consistency |
Background Consistency |
Motion Smoothness |
Imaging Quality |
Aesthetic Quality |
VQ | MQ | TA | |
| CausVid | 97.91 | 96.74 | 98.15 | 66.32 | 59.71 | -0.144 | 0.328 | 0.501 |
| Self-forcing | 97.18 | 96.37 | 98.35 | 68.35 | 59.19 | -0.461 | -0.216 | 0.656 |
| LongLive | 97.90 | 96.82 | 98.78 | 68.99 | 61.56 | -0.169 | -0.002 | 1.073 |
| Stream-T1 (on LongLive) | 98.43 | 97.18 | 99.03 | 69.10 | 62.11 | -0.073 | 0.226 | 1.170 |
| Relative gain | Δ0.54% | Δ0.37% | Δ0.25% | Δ0.16% | Δ0.89% | Δ56.8% | Δ11400% | Δ9% |
Ablation
To thoroughly validate the individual contributions of our proposed components, we conduct ablation studies on the 30s video. Both quantitative metrics and extensive qualitative visual comparisons consistently demonstrate the necessity and efficacy of our designed modules.
| Method | VBench Long↑ | VideoAlign↑ | ||||||
|---|---|---|---|---|---|---|---|---|
| Subject Consistency |
Background Consistency |
Motion Smoothness |
Imaging Quality |
Aesthetic Quality |
VQ | MQ | TA | |
| w/o Stream‑Scaled Memory Sinking | 98.30 | 97.04 | 98.92 | 69.51 | 61.90 | -0.083 | 0.188 | 1.146 |
| w/o Stream‑Scaled Noise Propagation | 98.35 | 97.14 | 98.98 | 69.07 | 61.99 | -0.094 | 0.176 | 1.164 |
| w/o Stream‑Scaled Reward Pruning | 98.04 | 96.88 | 98.87 | 69.17 | 61.22 | -0.173 | 0.014 | 1.035 |
| Ours | 98.43 | 97.18 | 99.03 | 69.10 | 62.11 | -0.073 | 0.226 | 1.170 |
BibTeX
@misc{tu2026streamt1testtimescalingstreaming,
title={Stream-T1: Test-Time Scaling for Streaming Video Generation},
author={Yijing Tu and Shaojin Wu and Mengqi Huang and Wenchuan Wang and Yuxin Wang and Chunxiao Liu and Zhendong Mao},
year={2026},
eprint={2605.04461},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.04461},
}