Research map / Video / Video generation
TinyHistory: Lightweight Video History Embeddings via Two-Stage Context Learning
lvmin zhang, Shengqu Cai, Muyang Li, Chong Zeng, Beijia Lu, Anyi Rao, Song Han, Gordon Wetzstein, Maneesh Agrawala
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 Stage I randomized frame-query pretraining (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char / FaceID: Char 52.64 -> 80.19; FaceID 58.38 -> 70.71 | Improved |
| train the encoder jointly in Stage II instead of freezing it (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char: 61.42 -> 80.19 | Improved |
| add the LR branch to the HR-only encoder (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char: 67.73 -> 80.19 | Improved |
| add the HR branch to the LR-only encoder (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char: 74.29 -> 80.19 | Improved |
| lower spatial/temporal compression (4x4x2 -> 2x2x2) (ablation) | TinyHistory · Storyboard generation | Scene / ELO: Scene 72.21 -> 87.11; ELO 1262 -> 1332 | Improved |
| halve context cost from 2x2x1 to 2x2x2 (ablation) | TinyHistory · Storyboard generation | ELO: 1342 -> 1332 | No clear effect |
| use 4x temporal instead of 4x spatial compression (2x2x4 vs 4x4x2) (ablation) | TinyHistory · Storyboard generation | Char: 80.19 -> 90.24 | Improved |
| add 3-frame sliding window overlap (ablation) | TinyHistory 4x4x2 · Storyboard generation | Scene: 72.21 -> 82.51 | Improved |
| add cross-attention from encoder to DiT (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char / Obj: Char 80.19 -> 87.01; Obj 74.38 -> 80.75 | Improved |
| add a second encoder (2x2x8) (ablation) | TinyHistory 4x4x2 · Storyboard generation | Char / Obj: Char 80.19 -> 88.25; Obj 74.38 -> 83.33 | Improved |
| cache history KV across diffusion steps (ablation) | TinyHistory 4x4x2 · Storyboard generation | FaceID / Scene: FaceID 70.71 -> 63.61; Scene 72.21 -> 65.28 | Hurt |
| weaker history conditioning (no pretraining or frozen encoder) (ablation) | TinyHistory · Storyboard generation | Dynamic degree: 84.31 -> 95.69 (no Stage I) | Mixed |
Closest papers by meaning
- Improving Autoregressive Video Modeling with History Understanding (ICLR 2026 · similarity 0.87)
- Learning-to-Memorize: Dynamic Context Management for Long-Horizon Autoregressive Video Generation (NeurIPS 2026 · similarity 0.87)
- History-Guided Video Diffusion (ICML 2025 · similarity 0.85)
- Seq2Time: Sequential Knowledge Transfer for Video LLM Temporal Grounding (CVPR 2025 · similarity 0.84)
- $\infty$-Video: A Training-Free Approach to Long Video Understanding via Continuous-Time Memory Consolidation (ICML 2025 · similarity 0.84)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.84)
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.
