Research map / Video / Video generation
Surprise Forcing: What to Remember, When to Skip in Long Video Generation
SHUWEI SHI, Zhen Li, Muyao Niu, Chuanhao Li, Bo Zheng, Kaipeng Zhang, zheng yinqiang
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 |
|---|---|---|---|
| use Budget-Norm gating instead of fixed threshold z=0.1 (ablation) | Surprise Forcing · VBench-Long (30s prompts) | Total score: 81.24 -> 83.20 | Improved |
| use Budget-Norm gating instead of fixed threshold z=2.0 (ablation) | Surprise Forcing · VBench-Long (30s prompts) | Total score: 81.66 -> 83.20 | Improved |
| use priority-based eviction instead of FIFO (ablation) | Surprise Forcing · VBench-Long (30s prompts) | Total score: 82.92 -> 83.20 | Improved |
| increase skip ratio from 0.4 to 0.6 (ablation) | Surprise Forcing · VBench-Long (30s prompts) | Total score: 83.20 -> 82.15 | Hurt |
| increase skip ratio from 0 to 0.4 (ablation) | Surprise Forcing · VBench-Long (30s prompts) | Total score: 84.03 -> 83.20 | Mixed |
| use similarity top-3 routing instead of recent-frame routing (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Subject consistency: 96.25 -> 97.16 | Improved |
| use similarity top-3 routing instead of random routing (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Subject consistency: 96.23 -> 97.16 | Improved |
| route 6 frames instead of 3 (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Dynamic degree: 56.56 -> 54.42 | Hurt |
| increase bank size from 6 to 12 (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | FPS: 17.18 -> 16.70 | Hurt |
| increase target write ratio from 0.3 to 0.1 (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Dynamic degree: 56.56 -> 55.16 | Hurt |
| increase target write ratio from 0.3 to 0.7 (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Dynamic degree: 56.56 -> 54.74 | Hurt |
| use priority-based eviction instead of greedy least-correlation eviction (ablation) | Surprise Forcing · VBench-Long (60s MovieGen) | Dynamic degree: 55.00 -> 56.56 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 8 of 9 shared results
- Rolling Forcing — Rolling Forcing: Autoregressive Long Video Diffusion in Real Time (ICLR 2026) · better on 3 of 3 shared results
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
- Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression (ICML 2026 · similarity 0.89)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.89)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.87)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.87)
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.
