Research map / Video / Video-language understanding
Make Your Training Flexible: Towards Deployment-Efficient Video Models
Chenting Wang, Kunchang Li, Tianxiang Jiang, Xiangyu Zeng, Yi Wang, Limin Wang
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 group-dynamic (L2, 4 groups) token selection instead of random mask (ablation) | Flux-Single · K400 w/ TO | top-1 accuracy: 79.0 -> 80.3 | Improved |
| use group-dynamic (L2, 4 groups) token selection instead of random mask (ablation) | Flux-Single · SSv2 w/ TO | top-1 accuracy: 65.9 -> 67.3 | Improved |
| add Flux tuning with token optimization to InternVideo2-S (ablation) | InternVideo2-S · K400, 512 tokens | top-1 accuracy: +7.5 | Improved |
| add Global-Local positional embedding (GLPE) (ablation) | Flux-Single + ViT · K400 w/ TO | top-1 accuracy: 80.3 -> 81.3 | Improved |
| add Dual Patch Normalization (ablation) | Flux-Single + ViT · K400 w/ TO | top-1 accuracy: 80.3 -> 81.2 | Improved |
| add both FluxViT modules (ablation) | Flux-Single · SSv2 w/ TO | top-1 accuracy: 67.3 -> 69.3 | Improved |
| replace LPE with RoPE (ablation) | Flux-Single + ViT · K400 w/ TO | top-1 accuracy: 81.0 -> 80.7 | Hurt |
| train with multiple token counts instead of single (ablation) | FluxViT · K400 w/ TO | top-1 accuracy: 81.7 -> 82.4 | Improved |
| multi-number pretraining and tuning with alignment (ablation) | FluxViT-S · K400, 512 test tokens | top-1 accuracy: 65.8 -> 76.6 | Improved |
| add Flux augmentation to pretraining and tuning (ablation) | FluxViT-S · K400, 512 test tokens | top-1 accuracy: 62.1 -> 65.8 | Improved |
| enlarge Tthres to (2048, 6144) (ablation) | Flux-Single FluxViT · K400 8x224 | top-1 accuracy: 80.5 -> 80.1 | Hurt |
| remove varied spatial resolution (ablation) | Flux-Single FluxViT · K400 with TO | top-1 accuracy: 81.7 -> 81.4 | No clear effect |
Closest papers by meaning
- PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference (ECCV 2026 · similarity 0.87)
- FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding (CVPR 2026 · similarity 0.86)
- Empowering World Models with Reflection for Embodied Video Prediction (ICML 2025 · similarity 0.85)
- FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.85)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.85)
- FPSAttention: Training-Aware FP8 and Sparsity Co-Design for Fast Video Diffusion (NeurIPS 2025 · similarity 0.85)
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.
