Ask or search anything...

GraphCanvas3D introduces a framework for controllable 3D scene generation using a hierarchical, graph-driven structure and iterative optimization guided by Multimodal Large Language Models. This framework achieves dynamic scene modification, including 4D temporal evolution, and surpasses existing text-to-3D methods with a CLIP score of 29.67 and an MLLM score of 8.3, alongside strong user preference in qualitative evaluations.