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

Qwen2-VL is an advanced upgrade of the Qwen-VL models that redefines the conventional predetermined-resolution approach to visual processing. Its Naive Dynamic Resolution mechanism dynamically converts images of varying resolutions into different numbers of visual tokens, yielding more efficient and accurate visual representations that align more closely with human perception. The model also integrates Multimodal Rotary Position Embedding (M-RoPE) to fuse positional information across text, images, and videos, and it handles both images and videos under a single unified processing paradigm.

To probe the potential of large multimodal models, the authors study scaling laws for large vision-language models, releasing versions at 2B, 8B, and 72B parameters and scaling the training data accordingly. The resulting series is highly competitive: the Qwen2-VL-72B model matches leading systems such as GPT-4o and Claude3.5-Sonnet across a variety of multimodal benchmarks while outperforming other generalist models, and the code is released publicly.