Research map / Video / Video generation
Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency
Zihan Su, Teng Hu, Jiangning Zhang, Ruiyan Wang, Ran Yi, Lizhuang Ma, Dacheng Tao
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 cycle-consistent learning (CCL) to baseline (ablation) | LongLive · Physical consistency 5s | PC: 61.94 -> 68.70 | Improved |
| add cycle-guided inference (CGI) to baseline (ablation) | LongLive · VBench 5s | Total score: 83.72 -> 84.51 | Improved |
| add CGI only (without CCL) (ablation) | LongLive · Physical consistency 5s | PC: 61.94 -> 65.67 | Mixed |
| add CGI on top of CCL (full Cycle-World) (ablation) | + CCL · Physical consistency 5s | PACE: 79.5 -> 81.91 | Improved |
| add CCL to baseline for 60s generation (ablation) | LongLive · Physical consistency 60s | PC: 59.45 -> 70.40 | Improved |
| add CCL to baseline for 60s generation (ablation) | LongLive · Physical consistency 60s | PACE: 75.71 -> 74.94 | Hurt |
| add CGI alone to baseline for 60s generation (ablation) | LongLive · Physical consistency 60s | PC: 59.45 -> 62.77 | Improved |
| raise cycle loss weight from 0.1 to 0.2 (ablation) | Cycle-World · VBench 5s | Total score: 84.05 -> 82.89 | Hurt |
| lower cycle loss weight from 0.1 to 0.05 (ablation) | Cycle-World · Physical consistency 5s | PC: 64.18 -> 56.72 | Hurt |
| spread 4 guidance steps on first timestep instead of 1 step at all 4 timesteps (ablation) | CGI · Physical consistency 5s | PACE: 81.91 -> 76.79 | Hurt |
| increase CGI step size from 10 to 15 (ablation) | CGI · VBench 5s | Total score: 84.36 -> 83.92 | Mixed |
| use Cycle-World instead of Self Forcing for 60s generation | VBench 60s · VBench 60s | Total score: 71.86 -> 82.88 | 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 6 of 7 shared results
- Rolling Forcing — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (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 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
- WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception (NeurIPS 2025 · similarity 0.90)
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (NeurIPS 2026 · similarity 0.88)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.88)
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.88)
- Learning World Models for Interactive Video Generation (NeurIPS 2025 · similarity 0.88)
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.88)
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.
