Chat-Edit-3D++: Interactive 3D and 4D Scene Editing via Large Language Models
CE3D++ turns language instructions into scheduled visual edits for 3D scenes and monocular 4D scenes.
The paper proposes a Hash-Atlas network that reframes 3D scene editing as operations on 2D atlas images. Its LLM-based editor accepts open-ended user text, interprets intent, and calls the needed visual models without a fixed input pattern. The 4D extension adds motion constraints for moving objects and uses a trajectory-editing dataset to help a smaller LLM schedule up to 30 visual tools. The authors say experiments show multi-round dialogue, scene understanding, and varied visual editing effects. HF Daily Papers' note
The paper proposes a Hash-Atlas network that reframes 3D scene editing as operations on 2D atlas images. Its LLM-based editor accepts open-ended user text, interprets intent, and calls the needed visual models without a fixed input pattern. The 4D extension adds motion constraints for moving objects and uses a trajectory-editing dataset to help a smaller LLM schedule up to 30 visual tools. The authors say experiments show multi-round dialogue, scene understanding, and varied visual editing effects. HF Daily Papers' note
score 4