Research map / 3D vision / 3D generation
Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Atlas3D adds a differentiable rigid-body simulation standability loss, a stable-equilibrium potential-energy loss, and Laplacian/normal regularizers to SDS-based text-to-3D pipelines so generated models stand stably in simulation and when 3D printed.
Problem
Text-to-3D models produced by SDS methods are visually plausible but often fall over in physics simulators or when 3D printed, lacking the physical stability needed for simulation, embodied AI, and fabrication.
Approach
A plug-in refinement stage that jointly optimizes mesh geometry with SDS loss and a differentiable-simulation standability loss (rotation deviation after settling), augmented by a gravitational-potential-based stable-equilibrium loss and smoothness regularizers, instead of post-hoc shape or topology optimization.
Key results
- Mean TRD reduced from 0.389 (Magic3D) to 0.060 (Atlas3D) over 107 prompts, more than six times lower.
- 3D-printed figures: 92.25% user-study success for Atlas3D vs 7% for baseline over 800 trials.
- Robot-arm placement: 90.6% success for Atlas3D vs 6.25% for baseline over 64 trials.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add differentiable physical constraints during Magic3D refinement | Magic3D · 107 prompts | TRD: 0.389 -> 0.060 | Helped |
| Add differentiable physical constraints during Magic3D refinement | Magic3D · 107 prompts | CLIP: 25.781 -> 25.356 | Hurt |
| Add differentiable physical constraints during Magic3D refinement | Magic3D · 107 prompts | Elo (GPT-4o): 1029.226 -> 970.774 | Hurt |
| Apply physical constraints during generation | Text-to-3D baseline · 3D-printed figures, user study (800 trials) | success rate: 7% -> 92.25% | Helped |
| Apply physical constraints during generation | Text-to-3D baseline · 3D-printed figures, robot arm (64 trials) | success rate: 6.25% -> 90.6% | Helped |
| Add stable equilibrium loss (ablation) | Atlas3D · Goose, perturbation angles 0, 0.01, 0.02, 0.04, and 0.08 | success rate: 0.97 -> 1; 0.71 -> 0.99; 0.62 -> 0.69; 0.23 -> 0.4 | Helped |
| Add standability loss (ablation) | Atlas3D · Simulation standability | standing success | Helped |
| Add stable equilibrium loss (ablation) | Atlas3D · Stability under perturbation | stability | Helped |
| Add bottom Laplacian regularization (ablation) | Atlas3D · Generated mesh geometry | surface smoothness | Helped |
| Integrate mesh regularization jointly with SDS instead of applying it as post-processing (ablation) | Atlas3D · Text alignment during refinement | text alignment | Helped |
Surprising findings
- Atlas3D's CLIP score (25.356) is slightly lower than the Magic3D baseline (25.781), and its GPT-4o Elo score is also lower (970.8 vs 1029.2).
- Post-processing the mesh with the physics losses (rather than joint optimization) degrades text alignment.
Limitations
- Optimizing all mesh vertices gives large freedom and can produce distorted meshes.
- Only SDS-based and text-to-3D settings are covered; generalization to non-SDS or image-to-3D is untested.
- Rigid-body assumption limits applicability to deformable objects.
Experimental setup
Datasets: text prompts sampled from a prior prompt set (107 after filtering)
Benchmarks: 107 text prompts (Magic3D baseline comparison)
Metrics: TRD (time-averaged rotation deviation) CLIP score Elo (GPT-4o) success rate under perturbation user study success rate robot arm success rate
Baselines: Magic3D MVDream make-it-stand
Builds on: DreamFusion Magic3D MVDream Warp Deep Marching Tetrahedra
Closest papers by meaning
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.86)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.86)
- PhysX-3D: Physical-Grounded 3D Asset Generation (NeurIPS 2025 · similarity 0.86)
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.85)
- Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image (NeurIPS 2026 · similarity 0.85)
- LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.