Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

By Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Ke-Yang Chen, Xuejing Liu, Jialin Wang, Wenbin Ge and 9 more. arXiv.org, 2024

4,557 citations recorded by the source corpus.

We present the Qwen2-VL Series, an advanced upgrade of the previous Qwen-VL models that redefines the conventional predetermined-resolution approach in visual processing. Qwen2-VL introduces the Naive Dynamic Resolution mechanism, which enables the model to dynamically process images of varying resolutions into different numbers of visual tokens. This approach allows the model to generate more efficient and accurate visual representations, closely aligning with human perceptual processes. The model also integrates Multimodal Rotary Position Embedding (M-RoPE), facilitating the effective fusion of positional information across text, images, and videos. We employ a unified paradigm for processing both images and videos, enhancing the model's visual perception capabilities. To explore the potential of large multimodal models, Qwen2-VL investigates the scaling laws for large vision-language models (LVLMs). By scaling both the model size-with versions at 2B, 8B, and 72B parameters-and the amount of training data, the Qwen2-VL Series achieves highly competitive performance. Notably, the Qwen2-VL-72B model achieves results comparable to leading models such as GPT-4o and Claude3.5-Sonnet across various multimodal benchmarks, outperforming other generalist models. Code is available at https://github.com/QwenLM/Qwen2-VL .

Subfields: Experimental and Cognitive Psychology

View Paper: https://www.semanticscholar.org/paper/a22b23b65cdbc0335b872ed7c7a7206c710f8997

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Experimental and Cognitive Psychology

1. Noah A. Smith
2. Zeerak Talat

Citations in the corpus

• InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency (2025, neutral): “Part of the results are collected from [138].”
• DanceGRPO: Unleashing GRPO on Visual Generation (2025, neutral): “Trajectories consist of T timesteps, after which P leads to a termination state. r ( z 0 , c ) is the reward model, which is always parametrized by a Vision-Language model (such as CLIP [26] and Qwen-VL [29]).”
• R1-VL: Learning to Reason with Multimodal Large Language Models via Step-Wise Group Relative Policy Optimization (2025, neutral): “In our experiments, we adopt two state-of-the-art open-source MLLMs, i.e., Qwen2-VL-2B and Qwen2-VL-7B [37].”
• Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs (2025, neutral): “…input modality, numerous vision-language models have been developed, including the LLava series [LLWL24, LLL + 24, LZG + 24], QWenVL series [BBY + 23, WBT + 24], InternVL series [CWT + 24, CWW + 24, CWC + 24], InternLM-XComposer series [ZDW + 23, DZZ + 24b], Molmo [DCL + 24], and NVLM [DLW + 24].”
• VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model (2025, neutral): “To-day, some of the most widely adopted open-source VLM series include LLaVA[25, 30, 31], QwenVL[6, 50], and In-ternVL [10, 11].”
• Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Multimodal Models (2024, background): “Broadly speaking, the academic benchmark results and human evaluation strongly agree, with the exception of Qwen2-VL [31], which performs strongly on the academic benchmarks and comparatively underperforms in the human evaluation.”
• Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces (2024, neutral): “…82] and the feature extraction abilities of modern vision encoders [30, 65, 69], MLLMs, especially visual MLLMs, exhibit unprecedented visual understanding capabilities [34, 40, 78, 86, 91, 102, 103], a promising direction toward developing world models [50] and embodied agents [17, 21, 37, 59].”
• Visual-RFT: Visual Reinforcement Fine-Tuning (2025, neutral): “Pioneered by LISA [13], there have been works done to enable large language models (LLMs) to output control tokens for other models (such as SAM [11]) or directly predict bounding box coordinates [31, 40] through supervised fine-tuning.”
• GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning (2025, positive): “Within the vision encoder, we adopt a strategy similar to Qwen2-VL [57], replacing the original 2D convolutions with 3D convolutions.”
• Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling (2024, neutral): “For example, Qwen2-VL [246] processes a cumulative total of 1.4 trillion tokens, whereas our InternVL2.”
• VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning (2025, neutral): “5-VL-72B [Wang et al., 2024b] scoring 74.8% and 57.2% on the same benchmarks.”
• LlaVA-CoT: Let Vision Language Models Reason Step-By-Step (2024, neutral): “Our results show that LLaVA-CoT consistently out-performs many open-source models of similar or even larger sizes, such as InternVL2-8B [11], Ovis1.”
• LVBench: An Extreme Long Video Understanding Benchmark (2024, neutral): “The rapid development of Multi-modal Large Language Models (MLLMs) has led to significant advances in vision-language understanding [26, 37, 38], driven by break-throughs in both large language models [3, 8, 25] and visual encoders [30, 35, 46].”
• InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models (2025, neutral): “Yet, the majority of leading MLLMs—both open-source and proprietary—are adapted from text-only large language models through sophisticated multi-stage pipelines [21, 19, 18, 5, 121, 7].”
• Emerging Properties in Unified Multimodal Pretraining (2025, positive): “Both MoE and MoT in our model use hard routing: the newly replicated generation expert exclusively processes VAE tokens, while the original parameters—the understanding expert—handle text and ViT tokens, following the strategy of the Qwen-VL series [4, 77].”