Research map / Video / Video generation
Towards Memory-Efficient Autoregressive Video Generation via Instance-Specific Parametric Absorption
Xiaomeng Fu, Jia Li, Yiming Hu, Yong Wang, Hayden So, Jiao Dai, Xiangxiang Chu, Jizhong Han
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 instance-specific weight absorption (Delta W) to local-attention layers (ablation) | ISPA (LongLive) · qualitative 30s generation | subject consistency | Improved |
| keep sink frame in local attention (ablation) | ISPA (LongLive) · qualitative 30s generation | visual sharpness | Improved |
| evict 50% of KV cache with ISPA-0.5 | LongLive 1.3B · VBench 5s (MovieGen prompts) | aesthetic quality: 64.43 -> 64.48 | No clear effect |
| evict 50% of KV cache with ISPA-0.5 | Krea-Realtime 14B · VBench 30s (MovieGen prompts) | aesthetic quality: 62.15 -> 63.93 | Improved |
| evict 50% of KV cache with ISPA-0.5 | Krea-Realtime 14B · VBench 5s (MovieGen prompts) | subject consistency: 94.90 -> 92.32 | Hurt |
| evict 50% of KV cache with ISPA-0.5 | LongLive 1.3B · VBench 30s (MovieGen prompts) | aesthetic quality: 63.59 -> 62.98 | Hurt |
| evict 50% of KV cache with ISPA-0.5 | Reward-Forcing 1.3B · VBench 30s (MovieGen prompts) | dynamic degree: 66.92 -> 65.09 | Hurt |
| raise L-Layer fraction beyond 0.5 (K=0.7N) (ablation) | ISPA (LongLive) · VBench 30s | overall quality | Hurt |
| use warmup T_warm=6 instead of 12 (ablation) | ISPA (LongLive) · VBench 30s | quality metrics | Hurt |
| increase warmup from 12 to 24 frames (ablation) | ISPA (LongLive) · VBench 30s | quality metrics | No clear effect |
| add W8A8 quantization to ISPA | ISPA · Inference speed | overall speedup: 1.35x -> 1.86x | Improved |
| add W8A8 quantization to ISPA | ISPA · LongLive generation | aesthetic and image quality | No clear effect |
Closest papers by meaning
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.89)
- Addressable Memory for Video World Models (NeurIPS 2026 · similarity 0.89)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.88)
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.88)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 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.
