vvangfaye/Urbex-4B
Urbex-4B is a 4 billion parameter vision-language model based on Qwen3-VL, developed by vvangfaye. It is specifically designed for agentic spatial grounding in city-scale 3D scenes, enabling natural-language descriptions to be grounded through active spatial search. The model utilizes landmark retrieval, BEV zooming, and oblique-view inspection to predict target boxes, making it highly specialized for urban spatial understanding tasks.
Loading preview...
Urbex Qwen3-VL 4B Overview
Urbex-4B is a specialized 4 billion parameter vision-language model, built upon the Qwen3-VL architecture, developed by vvangfaye. Its primary function is agentic spatial grounding within complex city-scale 3D environments. This model excels at interpreting natural-language descriptions and actively searching within urban scenes to identify and locate specified targets.
Key Capabilities
- Spatial Grounding: Accurately grounds natural-language descriptions to specific locations within 3D urban scenes.
- Active Spatial Search: Employs an intelligent search mechanism involving landmark retrieval, Bird's Eye View (BEV) zooming, and oblique-view inspection.
- Target Box Prediction: Capable of predicting precise target boxes based on spatial understanding and search results.
- City-Scale Understanding: Optimized for processing and understanding large-scale urban environments.
Good For
- Robotics and Autonomous Navigation: Ideal for agents requiring precise spatial understanding and object localization in urban settings.
- Geospatial AI Applications: Useful for tasks involving interpreting natural language queries about 3D city models.
- Urban Planning and Analysis: Can assist in identifying and analyzing features within digital twins of cities based on descriptive inputs.
- Research in Vision-Language Models: Provides a strong foundation for further research into spatial reasoning and agentic behavior in complex visual environments. More details can be found in the associated paper and codebase.