Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models

Benchmark Model Rank Results
visual-question-answering-vqa-on-hr-bench-4kCogVLM-Chat-17B–Overall: 45.5FSP: 49.5FCP: 41.5
visual-question-answering-vqa-on-hr-bench-4kDC² (InternVL-1.5-26B)–Overall: 63.4FSP: 73.3FCP: 53.5
visual-question-answering-vqa-on-hr-bench-4kDC² (LLaVA-v1.5-13B)–Overall: 51.5FSP: 52.0FCP: 51.0
visual-question-answering-vqa-on-hr-bench-4kDC² (LLaVA-v1.6-7B)–Overall: 50.0FSP: 53.0FCP: 47.0
visual-question-answering-vqa-on-hr-bench-4kDC² (Yi-VL-6B)–Overall: 44.0FSP: 44.0FCP: 44.0
visual-question-answering-vqa-on-hr-bench-4kDeepseek-VL-7B–Overall: 35.5FSP: 36.8FCP: 34.3
visual-question-answering-vqa-on-hr-bench-4kGPT-4o–Overall: 59.0FSP: 70.0FCP: 48.0
visual-question-answering-vqa-on-hr-bench-4kGemini 1.5 Flash–Overall: 66.8FSP: 76.8FCP: 56.8
visual-question-answering-vqa-on-hr-bench-4kInternLM-Xcomposer-7B–Overall: 39.3FSP: 37.3FCP: 41.3
visual-question-answering-vqa-on-hr-bench-4kInternLM-Xcomposer2-7B–Overall: 46.0FSP: 45.5FCP: 46.5
visual-question-answering-vqa-on-hr-bench-4kInternVL-1.5-26B–Overall: 60.6FSP: 69.5FCP: 51.8
visual-question-answering-vqa-on-hr-bench-4kInternVL-2-llama3-76B–Overall: 71.0FSP: 82.0FCP: 60.0
visual-question-answering-vqa-on-hr-bench-4kLLaVA-HR-X-13B–Overall: 53.6FSP: 61.3FCP: 46.0
visual-question-answering-vqa-on-hr-bench-4kLLaVA-HR-X-7B–Overall: 52.0FSP: 57.8FCP: 46.3
visual-question-answering-vqa-on-hr-bench-4kLLaVA-v1.5-13B–Overall: 43.3FSP: 45.2FCP: 41.3
visual-question-answering-vqa-on-hr-bench-4kLLaVA-v1.5-7B–Overall: 36.1FSP: 38.5FCP: 33.8
visual-question-answering-vqa-on-hr-bench-4kLLaVA-v1.6-13B–Overall: 45.5FSP: 49.8FCP: 41.3
visual-question-answering-vqa-on-hr-bench-4kLLaVA-v1.6-34B–Overall: 52.9FSP: 55.3FCP: 50.5
visual-question-answering-vqa-on-hr-bench-4kLLaVA-v1.6-7B–Overall: 47.9FSP: 49.0FCP: 46.8
visual-question-answering-vqa-on-hr-bench-4kMPLUG2-7B–Overall: 36.9FSP: 38.0FCP: 35.8
visual-question-answering-vqa-on-hr-bench-4kPhi3-Vision-4.2B–Overall: 48.1FSP: 54.3FCP: 42.0
visual-question-answering-vqa-on-hr-bench-4kQwen-VL-7B–Overall: 31.8FSP: 30.5FCP: 33.0
visual-question-answering-vqa-on-hr-bench-4kQwen-VL-Max–Overall: 58.5FSP: 65.0FCP: 52.0
visual-question-answering-vqa-on-hr-bench-4kQwen-VL-Plus–Overall: 53.0FSP: 65.0FCP: 41.0
visual-question-answering-vqa-on-hr-bench-4kSEAL-7B–Overall: 38.1FSP: 47.0FCP: 29.3
visual-question-answering-vqa-on-hr-bench-4kVisualGLM-6B–Overall: 20.5FSP: 22.5FCP: 18.5
visual-question-answering-vqa-on-hr-bench-4kXcomposer2-4kHD-7B–Overall: 57.8FSP: 63.8FCP: 51.8
visual-question-answering-vqa-on-hr-bench-4kYi-VL-34B–Overall: 44.4FSP: 46.0FCP: 42.8
visual-question-answering-vqa-on-hr-bench-4kYi-VL-6B–Overall: 42.6FSP: 42.8FCP: 42.5
visual-question-answering-vqa-on-hr-bench-4kfuyu-8B–Overall: 25.8FSP: 25.5FCP: 26.0
visual-question-answering-vqa-on-hr-bench-4kidefics-9B–Overall: 30.6FSP: 36.3FCP: 25.0
visual-question-answering-vqa-on-hr-bench-4kminigptv2-7B–Overall: 25.5FSP: 25.8FCP: 25.3
visual-question-answering-vqa-on-hr-bench-8kCogVLM-Chat-17B–Overall: 41.1FSP: 42.5FCP: 39.8
visual-question-answering-vqa-on-hr-bench-8kDC² (InternVL-1.5-26B)–Overall: 61.3FSP: 75.0FCP: 47.5
visual-question-answering-vqa-on-hr-bench-8kDC² (LLaVA-v1.5-13B)–Overall: 40.5FSP: 40.0FCP: 41.0
visual-question-answering-vqa-on-hr-bench-8kDC² (LLaVA-v1.6-7B)–Overall: 42.3FSP: 40.5FCP: 45.0
visual-question-answering-vqa-on-hr-bench-8kDC² (Yi-VL-6B)–Overall: 40.0FSP: 39.0FCP: 41.0
visual-question-answering-vqa-on-hr-bench-8kDeepseek-VL-7B–Overall: 33.4FSP: 33.8FCP: 33.0
visual-question-answering-vqa-on-hr-bench-8kGPT-4o–Overall: 55.5FSP: 62.0FCP: 49.0
visual-question-answering-vqa-on-hr-bench-8kGemini 1.5 Flash–Overall: 62.8FSP: 69.2FCP: 56.7
visual-question-answering-vqa-on-hr-bench-8kInternLM-Xcomposer-7B–Overall: 35.1FSP: 34.5FCP: 35.8
visual-question-answering-vqa-on-hr-bench-8kInternLM-Xcomposer2-7B–Overall: 37.9FSP: 36.0FCP: 39.8
visual-question-answering-vqa-on-hr-bench-8kInternVL-1.5-26B–Overall: 57.9FSP: 69.3FCP: 48.5
visual-question-answering-vqa-on-hr-bench-8kInternVL-2-llama3-76B–Overall: 61.4FSP: 69.0FCP: 53.8
visual-question-answering-vqa-on-hr-bench-8kLLaVA-HR-X-13B–Overall: 46.9FSP: 49.5FCP: 44.3
visual-question-answering-vqa-on-hr-bench-8kLLaVA-HR-X-7B–Overall: 41.6FSP: 42.0FCP: 41.3
visual-question-answering-vqa-on-hr-bench-8kLLaVA-v1.5-13B–Overall: 37.8FSP: 37.5FCP: 38.0
visual-question-answering-vqa-on-hr-bench-8kLLaVA-v1.5-7B–Overall: 32.1FSP: 33.0FCP: 31.3
visual-question-answering-vqa-on-hr-bench-8kLLaVA-v1.6-13B–Overall: 38.1FSP: 38.0FCP: 38.3
visual-question-answering-vqa-on-hr-bench-8kLLaVA-v1.6-34B–Overall: 47.4FSP: 44.5FCP: 50.3
visual-question-answering-vqa-on-hr-bench-8kLLaVA-v1.6-7B–Overall: 40.8FSP: 37.2FCP: 44.2
visual-question-answering-vqa-on-hr-bench-8kMPLUG2-7B–Overall: 33.8FSP: 33.8FCP: 33.8
visual-question-answering-vqa-on-hr-bench-8kPhi3-Vision-4.2B–Overall: 40.5FSP: 43.3FCP: 37.8
visual-question-answering-vqa-on-hr-bench-8kQwen-VL-7B–Overall: 28.4FSP: 28.8FCP: 28.0
visual-question-answering-vqa-on-hr-bench-8kQwen-VL-Max–Overall: 52.5FSP: 54.0FCP: 51.0
visual-question-answering-vqa-on-hr-bench-8kQwen-VL-Plus–Overall: 46.5FSP: 52.0FCP: 41.0
visual-question-answering-vqa-on-hr-bench-8kSEAL-7B–Overall: 35.6FSP: 42.5FCP: 28.8
visual-question-answering-vqa-on-hr-bench-8kVisualGLM-6B–Overall: 19.0FSP: 19.5FCP: 18.5
visual-question-answering-vqa-on-hr-bench-8kXcomposer2-4kHD-7B–Overall: 51.3FSP: 55.3FCP: 47.3
visual-question-answering-vqa-on-hr-bench-8kYi-VL-34B–Overall: 39.0FSP: 39.5FCP: 38.5
visual-question-answering-vqa-on-hr-bench-8kYi-VL-6B–Overall: 38.9FSP: 38.5FCP: 39.3
visual-question-answering-vqa-on-hr-bench-8kfuyu-8B–Overall: 26.1FSP: 24.8FCP: 27.5
visual-question-answering-vqa-on-hr-bench-8kidefics-9B–Overall: 28.8FSP: 34.8FCP: 22.8
visual-question-answering-vqa-on-hr-bench-8kminigptv2-7B–Overall: 26.1FSP: 26.0FCP: 26.3