Research map / Video / Video generation
LongLive: Real-time Interactive Long Video Generation
Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, Yuyang Zhao, Xianbang Wang, Muyang Li, Enze Xie, Ying-Cong Chen, Yao Lu, Song Han, Yukang Chen
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 KV recache at prompt switches instead of retaining the full KV cache (ablation) | LongLive · Prompt switch at 5s (10s video) | CLIP score: 25.92 -> 27.87 | Improved |
| add KV recache instead of clearing the cache at prompt switches (ablation) | LongLive · Prompt switch at 5s (10s video) | subject consistency: 89.59 -> 94.04 | Improved |
| KV recache instead of clearing cache (ablation) | LongLive · Prompt switch at 5s (10s video) | CLIP score: 28.95 -> 27.87 | Hurt |
| add short-window attention with frame sink | LongLive inference · single H100, long video | compute time and peak memory: -28% compute time, -17% peak memory | Improved |
| add frame sink to short-window attention (9 local + 3 sink) (ablation) | LongLive · 10s generation consistency | background and subject consistency | Improved |
| increase attention window from 3 to 27 latent frames (ablation) | LongLive · 10s generation consistency | consistency | Improved |
| train long (streaming long tuning) instead of train-short-test-long | LongLive vs Self-Forcing · VBench-Long 30s | total score: 81.59 -> 83.52 | Improved |
| increase LoRA rank from 32 to 256 (ablation) | streaming long tuning · VBench-Long 30s | total score: 81.08 -> 83.12 | Improved |
| increase LoRA rank from 256 to 512 (ablation) | streaming long tuning · VBench-Long 30s | total score: 83.12 -> 83.04 | No clear effect |
| full fine-tuning instead of LoRA rank 256 (ablation) | streaming long tuning · VBench-Long 30s | total score: 83.12 -> 83.52 | Improved |
| INT8 post-training quantization | LongLive · VBench | total score: 84.87 -> 84.31; 12.6 -> 16.4 FPS | Mixed |
| recaching cost at prompt switch during training | LongLive · 10s video with one switch | extra time cost: about 6% extra time | No clear effect |
Later papers that compared with this method
34 later accepted papers report LongLive as a baseline in their results tables; 33 of them report a better number on most of the results they share with it.
- Surprise Forcing: What to Remember, When to Skip in Long Video Generation (ECCV 2026 · better on 8 of 9 shared results)
- Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory (NeurIPS 2026 · better on 8 of 8 shared results)
- Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency (ECCV 2026 · better on 6 of 7 shared results)
- Anchor Forcing: Anchor Memory and Tri-Region RoPE for Interactive Streaming Video Diffusion (ECCV 2026 · better on 7 of 7 shared results)
- Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression (ICML 2026 · better on 5 of 6 shared results)
- Context Forcing: Consistent Autoregressive Video Generation with Long Context (ICML 2026 · better on 4 of 6 shared results)
- PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference (ECCV 2026 · better on 5 of 6 shared results)
- OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation (NeurIPS 2026 · better on 4 of 5 shared results)
26 more not listed.
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
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.90)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.89)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.89)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
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.
