From Pixels to Words -- Towards Native One-Vision Models at Scale
2026-05-27 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors explain that current vision-language models usually combine separate image and language parts in a way that loses fine detail across multiple images or frames. They introduce NEO-ov, a new model that processes multiple images or video frames all together in one system, keeping detailed connections between pixels and words. This method removes the need for separate modules, allowing better understanding of complex visual and language information over time. Their results show that NEO-ov matches or outperforms traditional models, proving this integrated approach works well at scale.
vision-language modelsimage encoderlanguage decoderspatiotemporal modelingpixel-word correspondencemulti-image understandingvideo understandingnative architecturemodular frameworkmultimodal modeling
Authors
Haiwen Diao, Jiahao Wang, Penghao Wu, Yuhao Dong, Yuwei Niu, Yue Zhu, Zhongang Cai, Weichen Fan, Linjun Dai, Silei Wu, Xuanyu Zheng, Mingxuan Li, Yuanhan Zhang, Bo Li, Hanming Deng, Huchuan Lu, Quan Wang, Lei Yang, Lewei Lu, Dahua Lin, Ziwei Liu
Abstract
Current vision-language models (VLMs) typically stitch together separate image encoders and language decoders via multi-stage alignment, a modular framework that inevitably fragments pixel-level signals across frames and scatters early pixel-word interactions. In parallel, native VLMs, despite impressive performance on single images, remain largely unexplored in multi-image, video understanding, and spatial intelligence. Hence, we introduce NEO-ov, a native foundation model that learns cross-frame and pixel-word correspondence end-to-end, without any external encoders, auxiliary adapters, or post-hoc fusion. By eliminating module boundaries entirely, NEO-ov enables fine-grained and unified spatiotemporal modeling to emerge natively inside the model. Notably, NEO-ov largely narrows the gap to modular counterparts while excelling at fine-grained visual perception, validating that native "one-vision" architectures are not only feasible but competitive at scale. Beyond empirical performance, we unveil systematic architectural analyses and detailed training recipes to facilitate subsequent native multimodal modeling. Our code and models are publicly available at: https://github.com/EvolvingLMMs-Lab/NEO.