Research map / Video / Video generation
LinVideo: A Post-Training Framework towards O(n) Attention in Efficient Video Generation
Yushi Huang, Xingtong Ge, Ruihao Gong, Chengtao Lv, Jun Zhang
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 learned selective transfer instead of manual layer choice (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 62.97 -> 66.07 | Improved |
| add learned selective transfer instead of heuristic grid search (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 60.74 -> 66.07 | Improved |
| add regularization Lreg (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 18.62 -> 66.07 | Improved |
| use ADM objective instead of MSE (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 61.56 -> 66.07 | Improved |
| use ADM instead of few-step DMD loss (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 57.44 -> 66.07 | Improved |
| estimate score with the model itself instead of an extra score model (ablation) | LinVideo · VBench (5 dims) | Overall Consistency: 25.87 -> 26.52 | Improved |
| self-score estimation reduces training time (ablation) | LinVideo · training cost | training time: ~4.4x less | Improved |
| increase target linear layers from 16 to 20 (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 66.07 -> 64.38; latency 68.26 -> 61.68 | Hurt |
| increase target linear layers from 10 to 16 (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 66.32 -> 66.07 | No clear effect |
| vary coefficient of Lcon+Lreg from 0.1 to 0.001 (ablation) | LinVideo · VBench (5 dims) | Imaging Quality: 66.21 / 65.98 vs 66.07 | No clear effect |
| apply LinVideo versus FlashAttention2 | Wan 1.3B · VBench-2.0 | total score: 56.74 vs 56.74 | No clear effect |
| apply LinVideo versus FlashAttention2 | Wan 14B · VBench-2.0 | total score: 59.85 -> 59.62 | Hurt |
Closest papers by meaning
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.88)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.88)
- Faster Video Diffusion with Trainable Sparse Attention (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.
