Stream-T1: Test-Time Scaling for Streaming Video Generation
Yijing Tu, Shaojin Wu, Mengqi Huang, Wenchuan Wang, Yuxin Wang, Chunxiao Liu, Zhendong Mao
arXiv:2605.04461 · 109▲ · cs.CV
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
Stream-T1 is a test-time scaling framework built on top of a streaming (chunk-by-chunk autoregressive) video diffusion model (LongLive). It combines three inference-time tricks: initializing each chunk's noise by spherical interpolation from the previous chunk's noise, beam search pruning of chunk candidates using a combined image-reward plus video-reward score, and reward-guided routing of evicted KV cache entries into discard, EMA-merge, or append pathways.
Why it matters
Applying test-time scaling to whole-video diffusion is expensive because search happens in a global high-dimensional space with many denoising steps. Doing it at the chunk level (4 denoising steps per chunk) makes candidate search cheaper and lets you correct temporal artifacts locally instead of discarding a whole video, which matters for generating long (30s) videos that stay consistent over time.
Practical takeaway
If you run a streaming video generator, you can bolt on candidate search plus reward-guided KV cache management at inference without retraining to reduce drift over long clips. Watch for these chunk-level TTS ideas showing up as inference options in streaming video tooling.
Key result
On 30s video generation (128 MovieGen prompts, evaluated with VBench-Long and VideoAlign), Stream-T1 beats its LongLive baseline on nearly all metrics, with the largest gains on the human-preference VideoAlign scores; note these are near the noise floor (motion quality went from -0.002 to 0.226, which the paper reports as a large percentage but off a tiny base). Gains on standard VBench consistency metrics are under 1 percent. It also outperforms Best-of-N and beam-search baselines applied to the same model.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.