Research map / Video / Video-language understanding
DIVE-Bench: Codec-Style High-FPS Video Understanding
Haichao Zhang, Wenhao Chai, Shwai He, Ang Li, Yun Fu
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 gated inter-tokenization to the baseline tokenizer (ablation) | LLaVA-OV 0.5B · DIVE-Bench educational | MOS: 1.66 -> 1.93 | Improved |
| add semantic-scene merging on top of the gated tokenizer (ablation) | GRT · DIVE-Bench educational | MOS: 1.93 -> 1.94 | No clear effect |
| replace LLaVA-OV tokenization with GRT | LLaVA-OV 0.5B · DIVE-Bench educational | MOS: 2.01 -> 2.50 | Improved |
| replace LLaVA-OV tokenization with GRT | LLaVA-OV · DIVE-Bench high-motion | Effective FPS: 4.49 -> 8.69 | Improved |
| replace LLaVA-OV tokenization with GRT | LLaVA-OV · DIVE-Bench high-motion | Grid ADE: 1.3618 -> 1.1456 | Improved |
| replace LLaVA-OV tokenization with GRT | LLaVA-OV · DIVE-Bench high-motion | Grid FDE: 1.6085 -> 1.1828 | Improved |
| apply GRT at matched low-FPS sampling | LLaVA-One Vision 0.5B · MLVU | score: 33.002 -> 34.066 | Improved |
| apply GRT at matched low-FPS sampling | LLaVA-One Vision 0.5B · VideoMME | score: 43.963 -> 44.037 | No clear effect |
| apply GRT on MLVU Anomaly Recognition | LLaVA-OV 0.5B · MLVU Anomaly Recognition | score: 30.800 -> 38.500 | Improved |
| apply GRT on MLVU Ego-centric | LLaVA-OV 0.5B · MLVU Ego-centric | score: 37.700 -> 39.600 | Improved |
| apply GRT on MLVU TutorialQA and Topic Reasoning | LLaVA-OV 0.5B · MLVU TutorialQA / Topic Reasoning | score: 27.900 -> 27.900; 62.600 -> 62.600 | No clear effect |
| use GRT gated tokenizer at 1 FPS | LLaVA-OV 0.5B · Tokenization time at 1 FPS | time (s): 0.0487 s -> 0.0226 s | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.91)
- Improving LLM Video Understanding with 16 Frames Per Second (ICML 2025 · similarity 0.89)
- Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.89)
- Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (CVPR 2025 · similarity 0.89)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.89)
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.
