Existing text-to-3D generation methods often neglect physical constraints, resulting in models that fail to stand independently. Atlas3D addresses this issue by enhancing existing Score Distillation Sampling (SDS)-based tools to generate self-supporting 3D models that adhere to physical laws of stability, contact, and friction.
Atlas3D introduces a novel differentiable simulation-based loss function combined with physically inspired regularization. This method can serve as a refinement or post-processing module for existing 3D generation frameworks. It ensures the models produced are stable and can stand without additional supports, which is crucial for applications in interactive gaming, embodied AI, and robotics.
The approach integrates physics-based guidance into the 3D generation process, particularly focusing on the stability of man-made objects like action figures and toys. By incorporating a differentiable rigid-body simulator, Atlas3D optimizes both the geometry and physical attributes of the generated models to achieve standability and stable equilibrium. Additionally, geometry regularization enhances the smoothness of the meshes.
Atlas3D has been validated through extensive generation tasks and real-world tests. The resulting models demonstrate improved stability compared to those produced by existing methods. These stable models can be directly used in physical simulators or 3D printed for real-world applications, significantly reducing the need for manual adjustments and supports.
You can read the full research paper, titled “Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication” over at this link.
