Research map / Vision-language and multimodal / Multimodal LLMs
HRScene: How Far Are VLMs from Effective High-Resolution Image Understanding?
Yusen Zhang, Wenliang Zheng, Aashrith Madasu, Peng Shi, Ryo Kamoi, Hao Zhou, Zhuoyang Zou, Shu Zhao, Sarkar Snigdha Sarathi Das, Vipul Gupta, Xiaoxin Lu, Nan Zhang and 5 more
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 index text under each sub-image (ablation) | ComplexGrid construction (Qwen2-VL 7B) · ComplexGrid 10x10 | accuracy: 1.53 -> 8.22 | Improved |
| add index text under each sub-image (ablation) | ComplexGrid construction (InternVL2 8B) · ComplexGrid 10x10 | accuracy: 2.78 -> 17.15 | Improved |
| add white interval between sub-images (ablation) | ComplexGrid construction (Qwen2-VL 72B) · ComplexGrid 10x10 | accuracy: 3.17 -> 5.17 | Improved |
| add both index and interval to dense grid (ablation) | ComplexGrid construction (InternVL2 8B) · ComplexGrid 10x10 | accuracy: 2.78 -> 20.54 | Improved |
| add both index and interval to dense grid (ablation) | ComplexGrid construction (Qwen2-VL 72B) · ComplexGrid 10x10 | accuracy: 3.17 -> 18.72 | Improved |
| scale model size within a family | VLMs · HRScene real-world | accuracy | Improved |
| increase whitebackground grid from 1x1 to 10x10 | DeepSeek-VL2-27B · WhiteBackground NIAH | performance: 72.06 -> 23.95 | Hurt |
| increase whitebackground grid from 1x1 to 10x10 | Qwen2-VL - 72B · WhiteBackground NIAH | performance: 84.13 -> 84.56 | No clear effect |
| use native-resolution input (Qwen) instead of size-limited input (GPT-4o) | Qwen 72B vs GPT 4o · Galaxy (20k resolution) | accuracy: 68.68 -> 80.80 | Improved |
| native-resolution input on 1k images | Qwen 72B vs GPT 4o · HRVQA (1k resolution) | accuracy: 73.82 -> 69.59 | Hurt |
| resize image before input (ablation) | Qwen2-VL 7B · Autonomous Driving and HR-Bench | accuracy | Mixed |
Closest papers by meaning
- Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.86)
- XLRS-Bench: Could Your Multimodal LLMs Understand Extremely Large Ultra-High-Resolution Remote Sensing Imagery? (CVPR 2025 · similarity 0.85)
- InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD (NeurIPS 2024 · similarity 0.85)
- VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning? (ICLR 2026 · similarity 0.84)
- Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.84)
- HRVDA: High-Resolution Visual Document Assistant (CVPR 2024 · 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.
