Research map / Video / Video generation
Fastcar: Cache Attentive Replay for Fast Auto-Regressive Video Generation on the Edge
Xuan Shen, Weize Ma, Yufa Zhou, Enhao Tang, Yanyue Xie, Zhengang Li, Yifan Gong, Quanyi Wang, Henghui Ding, Yiwei Wang, Pu Zhao, Jun Lin and 1 more
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 consistent instead of layer-wise inconsistent thresholds at 10% replay (ablation) | FastCar · VBench prompts (10% replay) | LPIPS: 32.94 -> 27.31 | Improved |
| use consistent instead of layer-wise inconsistent thresholds at 10% replay (ablation) | FastCar · VBench prompts (10% replay) | VBench total: 71.8% -> 73.4% | Improved |
| lower threshold tau below -2.5 (higher replay ratio) (ablation) | FastCar · VBench prompts | VBench total: 71.4% -> 71.5% | No clear effect |
| increase replay ratio from 10% to 80% | FastCar · VBench prompts | Latency (s): 629.1 -> 390.7 | Improved |
| increase replay ratio from 10% to 80% | FastCar · VBench prompts | VBench total: 73.4% -> 71.5% | Hurt |
| replace sparse attention (local 128) with FastCar at 80% replay | FastCar vs StreamingLLM · VBench prompts | VBench total: 60.7% -> 71.5% | Improved |
| shrink sparse attention local size from 256 to 16 | StreamingLLM · VBench prompts | Latency (s): 670.5 -> 662.7 | No clear effect |
| shrink sparse attention local size from 256 to 128 | StreamingLLM · VBench prompts | VBench total: 72.1% -> 60.7% | Hurt |
| add FastCar (87% replay) to sparse attention local 16 | StreamingLLM local 16 · VBench prompts | VBench total: 64.5% -> 71.6% | Improved |
| add FastCar (87% replay) to sparse attention local 16 | StreamingLLM local 16 · VBench prompts | Latency (s): 662.7 -> 324.31 | Improved |
| raise replay in FastCar+SA (local 256) from tau=-1 to tau=-4 | FastCar + sparse attention local 256 · VBench prompts | VBench total: 67.7% -> 71.8% | Improved |
Closest papers by meaning
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.88)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.87)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.87)
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.87)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.86)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.86)
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.
