MVRD: Multi-View Relational Distillation for Spatial Reasoning

This repository contains the model presented in the paper Multi-View Relational Distillation for Spatial Reasoning with Vision-Language Models. It is built on LLaVA-Video-7B-Qwen2 and includes additional grounding tokens for spatial reasoning tasks such as visual spatial understanding and 3D scene understanding.

For training and evaluation instructions, please refer to the GitHub repository.