Research map / Video / Video-language understanding
FLoC: Facility Location-Based Efficient Visual Token Compression for Long Video Understanding
Janghoon Cho, Jungsoo Lee, Munawar Hayat, Kyuwoong Hwang, Fatih Porikli, Sungha Choi
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 |
|---|---|---|---|
| compress with FLoC instead of DivPrune (1/32) | Qwen2.5-VL-7B · Avg of 4 benchmarks | accuracy: 55.89 -> 57.45 | Improved |
| use FLoC with 7200 frames compressed vs 768 frames full tokens | Qwen2.5-VL-7B · Avg of Video-MME/MLVU/LVB | accuracy: 65.82 -> 67.03 | Improved |
| use FLoC with 7200 frames vs 768 frames (32B) | Qwen2.5-VL-32B · Avg of Video-MME/MLVU/LVB | accuracy: 68.19 -> 70.63 | Improved |
| facility location instead of K-Means | Qwen2-VL-7B · Avg accuracy | accuracy: 58.66 -> 59.74 | Improved |
| facility location instead of K-Means (compression time) | Qwen2-VL-7B · 1/32 ratio T=32 | seconds: 59.00 -> 0.413 | Improved |
| FLoC vs Spectral Clustering on Video-MME overall | LLaVA-Video-7B · Video-MME 1/8 | accuracy: 61.81 -> 61.04 | Hurt |
| FLoC vs Pooling on Action Count | Qwen2-VL-7B · MLVU AC 1/32 | accuracy: 32.52 -> 23.3 | Hurt |
| FLoC vs DivPrune on Needle QA | Qwen2-VL-7B · MLVU NQA 1/32 | accuracy: 68.45 -> 71.27 | Improved |
| increase block length T beyond 4 (ablation) | FLoC · Qwen2-VL-7B block length | accuracy | Improved |
| FLoC vs DivPrune at 1/32 on InternVL3 Video-MME | InternVL3-8B · Video-MME 1/32 | accuracy: 60.85 -> 60.81 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FastVID — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 3 of 3 shared results
- LongVU — LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.92)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.91)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.90)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.90)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.90)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · 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.
