internlm/Spatial-SSRL-3B
Spatial-SSRL-3B is a 3 billion parameter vision-language model developed by InternLM, built upon Qwen2.5-VL-3B. It is specifically optimized for spatial understanding tasks through the application of Spatial-SSRL, a lightweight, self-supervised reinforcement learning paradigm. This model excels at interpreting spatial relationships in both image and video contexts while maintaining general visual capabilities, making it suitable for applications requiring advanced spatial intelligence.
Loading preview...
Spatial-SSRL-3B: Enhanced Spatial Understanding
Spatial-SSRL-3B is a 3 billion parameter vision-language model from InternLM, designed to significantly improve spatial understanding capabilities. Built on the Qwen2.5-VL-3B base, this model leverages a novel, lightweight self-supervised reinforcement learning (Spatial-SSRL) paradigm to achieve superior performance in spatial intelligence tasks.
Key Capabilities
- Advanced Spatial Intelligence: Excels in interpreting spatial relationships within images and videos, demonstrated across seven spatial understanding benchmarks.
- Lightweight and Tool-Free: Spatial-SSRL is a cost-effective framework that avoids reliance on human labels, API calls, or external tool annotations, reducing inherent errors and additional costs.
- Highly Scalable Training: Utilizes ordinary raw RGB and RGB-D images for data curation, enabling highly scalable training without richly-annotated public datasets.
- Preserves General Visual Abilities: Enhances spatial understanding while maintaining the original general visual capabilities of its base model.
Good For
- Applications requiring precise spatial reasoning from visual data.
- Research and development in self-supervised learning for vision-language models.
- Tasks benefiting from a cost-effective and scalable approach to improving spatial understanding in LLMs.