Research map / Video / Video generation
Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusion
Bowen Xue, Brandon Feng, Chenguo Lin, Yuchen Lin, Yujia Zeng, lvmin zhang, Maneesh Agrawala, Honglei Yan, Panwang Pan
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 sparse RoPE mapping compressed tokens to physical coordinates instead of index-based RoPE (ablation) | Ring Forcing · reconstruction | PSNR: 19.05 -> 24.22 | Improved |
| use sparse RoPE instead of standard RoPE (ablation) | Ring Forcing · reconstruction | LPIPS: 0.27 -> 0.13 | Improved |
| increase ring context drop probability from 0.0 to 1.0 (ablation) | Ring Forcing · ring-mode reconstruction | PSNR: 25.02 -> 21.86 | Hurt |
| increase ring context drop probability from 0.0 to 1.0 (ablation) | Ring Forcing · standard-mode diversity | Diversity Score: 3.71 -> 4.23 | Improved |
| choose p_drop=0.4 over 0.0 (ablation) | Ring Forcing · reconstruction vs diversity | PSNR / Diversity: PSNR 25.02 -> 24.22; Diversity 3.71 -> 4.11 | Mixed |
| timestep composition (global + detail streams) vs single-stream baselines (ablation) | Ring Forcing · reconstruction SSIM | SSIM | Improved |
| global stream only (high spatial compression) vs detail stream only (aggressive temporal compression) (ablation) | Ring Forcing · reconstruction SSIM | SSIM | Improved |
| ring-structured training vs linear-trained baselines at 1s gap | Ring Forcing · A-D-R benchmark | Texture consistency: 0.252 -> 0.723 | Improved |
| ring-structured training vs baselines at 1s gap | Ring Forcing · A-D-R benchmark | Geometry consistency: 0.368 -> 0.839 | Improved |
| Ring Forcing vs base model on aesthetics | Ring Forcing · General benchmark | Aesthetic: 5.825 -> 5.822 | No clear effect |
| longer disappearance gap (15s to 60s) | Ring Forcing · A-D-R benchmark | Texture consistency: 0.509 -> 0.396 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 2 of 2 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
- Learning-to-Memorize: Dynamic Context Management for Long-Horizon Autoregressive Video Generation (NeurIPS 2026 · similarity 0.91)
- Recurrent Autoregressive Diffusion: Global Memory Meets Local Attention (ECCV 2026 · similarity 0.90)
- Addressable Memory for Video World Models (NeurIPS 2026 · similarity 0.89)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.88)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- History-Guided Video Diffusion (ICML 2025 · 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.
