Friday, 23 October 2026 · 6 min read
WeVisOne-VL unifies object- and pixel-level vision-language tasks with structured token prediction
WeVisOne-VL integrates a lightweight pixel decoder into a vision-language model, enabling unified object detection, referring localization, counting, OCR, segmentation, and matting. Trained with curriculum fine-tuning and reinforcement learning, it leads on over 30 benchmarks for both discrete and dense visual tasks.
Based on the paper WeVisOne-VL: Unified Visual Perception from Objects to Pixels by Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu and 4 others
Challenges in unifying fine-grained visual perception with language
Vision-language models (VLMs) aim to understand images guided by natural language, but existing models struggle with fine-grained perception tasks like precise localization and pixel-level segmentation.
Traditional object detectors excel at dense detection but lack open-world recognition, while VLMs recognize broad concepts but often hallucinate or miss small or crowded objects.
A unified model that can handle both object-level and pixel-level tasks with robust rejection of absent targets is needed for practical deployment in applications like content creation and autonomous systems.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- The figure from the paper, and how to read it
- How it was tested, what was found, and where it falls short