Research map / Video / Video generation
MemRoPE: Training-Free Infinite Video Generation via Evolving Memory Tokens
Youngrae Kim, Qixin Hu, C.-C. Jay Kuo, Peter Beerel
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 MemRoPE dual-EMA memory tokens and Online RoPE Indexing (training-free) | Self-Forcing · VBench-Long 240 s | average score: 83.04 -> 84.89 | Improved |
| add MemRoPE | LongLive · VBench-Long 1 hour | average score: 87.01 -> 87.99 (vs infty-RoPE) | Improved |
| add Deep Forcing participative compression | Self-Forcing · VBench-Long 240 s | average score: 83.04 -> 81.66 | Hurt |
| add infty-RoPE | Self-Forcing · VBench-Long 120 s | average score: 84.25 -> 83.64 | Hurt |
| add Deep Forcing | LongLive · VBench-Long 240 s | average score: 85.33 -> 85.48 | No clear effect |
| add any memory stream (long, short or both) vs no memory (ablation) | MemRoPE on LongLive · VBench-Long 60 s | average score: 85.48 -> 85.97 | Improved |
| add long-term memory to short-term memory (ablation) | MemRoPE on LongLive · VBench-Long 60 s | average score: 85.95 -> 85.97 | Improved |
| store keys without RoPE before EMA aggregation instead of aggregating RoPE-rotated keys (ablation) | MemRoPE · VBench-Long 60 s Self-Forcing | average score: 85.28 -> 85.41 | Improved |
| store keys without RoPE before EMA aggregation instead of rotated keys (ablation) | MemRoPE · VBench-Long 60 s LongLive | average score: 86.08 -> 86.24 | Improved |
| vary EMA decay rates alpha_L and alpha_S across 12 combinations (ablation) | MemRoPE · VBench-Long | average score: <0.7 variation | No clear effect |
| use compact three-tier cache (C=12) in place of sliding window (C=21) for Self-Forcing | MemRoPE · inference speed | video FPS: 3.631 -> 5.115 | Improved |
| add MemRoPE to LongLive at equal cache size | LongLive · inference speed | video FPS: 4.409 -> 4.376 | No clear effect |
Closest papers by meaning
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.90)
- Anchor Forcing: Anchor Memory and Tri-Region RoPE for Interactive Streaming Video Diffusion (ECCV 2026 · similarity 0.89)
- Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression (ICML 2026 · similarity 0.89)
- FAST-AR: Fast Autoregressive Video Diffusion and World Models with Temporal Cache Compression and Sparse Attention (ICML 2026 · similarity 0.89)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (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.
