Research map / Large language models / Alignment and preferences
DSO: Aligning 3D Generators with Simulation Feedback for Physical Soundness
Ruining Li, Chuanxia Zheng, Christian Rupprecht, Andrea Vedaldi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DSO with DRO loss to TRELLIS | TRELLIS · Stability eval full set | % Stable: 85.1 -> 99.0 | Improved |
| add DSO with DPO loss to TRELLIS | TRELLIS · Stability eval full set | % Stable: 85.1 -> 95.1 | Improved |
| use DRO instead of DPO objective (ablation) | TRELLIS + DSO · Stability eval full set | % Stable: 95.1 -> 99.0 | Improved |
| add DSO (DRO) to TRELLIS on hard unstable subset | TRELLIS · Partial set (11 unstable objects) | % Stable: 54.5 -> 95.5 | Improved |
| add DSO (DRO) rather than test-time PhysComp | TRELLIS · Partial set (11 unstable objects) | Chamfer Distance: 0.0698 -> 0.0520 | Improved |
| add DSO (DRO) to TRELLIS | TRELLIS · Full set geometry | Chamfer Distance: 0.0485 -> 0.0440 | Improved |
| fine-tune on stable samples only (SFT) instead of DSO DRO (ablation) | TRELLIS · Full set | % Stable: 99.0 -> 89.5 | Hurt |
| train DRO on synthetic FLUX images instead of Objaverse renderings (ablation) | TRELLIS + DSO (DRO) · Full set | % Stable: 99.0 -> 97.6 | Hurt |
| train DPO on synthetic FLUX images instead of Objaverse renderings (ablation) | TRELLIS + DSO (DPO) · Full set | % Stable: 95.1 -> 93.5 | Hurt |
| train DRO 5000 steps instead of 4000 (ablation) | TRELLIS + DSO (DRO) · Full set | Chamfer Distance: 0.0440 -> 0.0853 | Hurt |
| train DRO 2000 steps instead of 4000 (ablation) | TRELLIS + DSO (DRO) · Full set | % Stable: 99.0 -> 91.5 | Hurt |
| replace DSO with flat-cut post-processing at z=0.2 (ablation) | TRELLIS · Full set | % Stable: 99.0 -> 95.8 | Hurt |
Closest papers by meaning
- Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication (NeurIPS 2024 · similarity 0.88)
- Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image (NeurIPS 2026 · similarity 0.86)
- PhysX-3D: Physical-Grounded 3D Asset Generation (NeurIPS 2025 · similarity 0.86)
- DecompDreamer: A Composition-Aware Curriculum for Structured 3D Asset Generation (NeurIPS 2026 · similarity 0.85)
- Interactive3D: Create What You Want by Interactive 3D Generation (CVPR 2024 · similarity 0.85)
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (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.
