Research map / Video / Video generation
UniTemp: Unlocking Video Generation in Any Temporal Order via Autoregressive Distillation
Lin Zhang, Sicheng Mo, Zefan Cai, Jinhong Lin, Zihao Lin, Jiuxiang Gu, Krishna Kumar Singh, Yuheng Li, Yin Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 3 blockwise anchor latents in backward generation (ablation) | UniTemp backward · Flickering ratio inter-block | inter-block FR: 1.42 -> 1.07 | Improved |
| add 1 anchor latent (ablation) | UniTemp backward · Flickering ratio inter-block | inter-block FR: 1.42 -> 1.26 | Improved |
| add 2 anchor latents (ablation) | UniTemp backward · Flickering ratio inter-block | inter-block FR: 1.42 -> 1.23 | Improved |
| include anchor latents as output at inference (ablation) | UniTemp backward P=3 · Flickering ratio inter-block | inter-block FR: 1.07 -> 1.46 | Hurt |
| include anchor latents as output in training and inference (ablation) | UniTemp backward P=3 · Flickering ratio inter-block | inter-block FR: 1.08 -> 1.48 | Hurt |
| share generator and fake critic across directions (ablation) | UniTemp forward · VBench short 5s | VBench Total: 83.42 -> 83.81 | Improved |
| share generator and fake critic across directions (ablation) | UniTemp backward · VBench short 5s | VBench Total: 84.26 -> 84.23 | No clear effect |
| share generator only (ablation) | UniTemp backward · VBench short 5s | VBench Total: 84.26 -> 83.96 | Hurt |
| anchor latents vs none for backward generation, user preference (ablation) | UniTemp backward · User study A/B (200 responses) | preference rate: 87% vs 13% | Improved |
| add future sink latents to forward long generation (ablation) | UniTemp forward · Long video 100s | dynamic degree: 87.50 -> 44.50 | Hurt |
| add future sink latents to forward long generation (ablation) | UniTemp forward · Long video 100s | subject consistency: 93.95 -> 97.30 | Improved |
| add past sink latents to backward generation with future sink (ablation) | UniTemp backward · Long video 100s | subject consistency: 93.94 -> 96.75 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 2 of 2 shared results
- Infinity-Rope — Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout (CVPR 2026) · better on 2 of 2 shared results
- Rolling-Forcing — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (ICLR 2026) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.92)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.92)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.91)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.91)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.90)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
