Research map / Image generation / Fast diffusion sampling
Walking the Schrödinger Bridge: A Direct Trajectory for Text-to-3D Generation
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.
Proposes TraCe, a text-to-3D method that reinterprets SDS as a special case of the Schrödinger Bridge and builds a diffusion bridge between the current rendering and a one-step denoised text-conditioned target, training a LoRA-adapted T2I model on it for 3D Gaussian Splatting optimization.
Problem
SDS-based text-to-3D generation requires high classifier-free guidance and produces artifacts such as over-saturation and over-smoothing, while noisy score estimates may not give an optimal optimization direction.
Approach
Treat 3D optimization as learning a direct diffusion bridge from the current rendering to a text-conditioned denoised target, using a LoRA-adapted diffusion model trained on this bridge's score dynamics, and show that SDS is a special case of a Schrödinger Bridge.
Key results
- TraCe reaches a CLIP Score (ViT-L/14) of 69.26% vs 68.61% for SDS across 83 prompts.
- TraCe achieves the highest GPTEval3D overall score of 1028.03 and ImageReward of -0.2855, vs SDS at 1018.09 and -0.4329.
- Ablation on ImageReward: full method -0.2486 vs -0.4488 with both LoRA and scheduled t-sampling off.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TraCe instead of SDS | text-to-3D generation with 3DGS · Dreamfusion gallery prompts (83 prompts, 120 views) | GPTEval3D Overall: 1018.09 -> 1028.03 | Helped |
| use TraCe instead of ISM | text-to-3D generation with 3DGS · Dreamfusion gallery prompts (83 prompts, 120 views) | CLIP Score ViT-L/14 (%): 69.0093$\pm$10.2400 -> 69.2609$\pm$7.8366 | Helped |
| use TraCe instead of ISM | text-to-3D generation with 3DGS · Dreamfusion gallery prompts (83 prompts, 120 views) | ImageReward: -0.3904$\pm$0.9503 -> -0.2855$\pm$0.8909 | Helped |
| add LoRA adaptation (ablation) | TraCe · Fox generation | ImageReward: -0.3389 $\pm$ 0.9721 -> -0.2486 $\pm$ 0.8909 | Helped |
| add scheduled t-sampling (ablation) | TraCe · Fox generation | ImageReward: -0.4020 $\pm$ 1.0019 -> -0.2486 $\pm$ 0.8909 | Helped |
| use TraCe instead of SDS | NeRF text-to-3D generation · NeRF evaluation | GPTEval3D Overall: 1000.00 -> 1179.86 | Helped |
| use TraCe instead of VSD | NeRF text-to-3D generation · NeRF evaluation | ImageReward: -0.6737 $\pm$ 0.4098 -> -0.4533 $\pm$ 0.4635 | Helped |
Limitations
- Runtime of about 14 minutes per asset, longer than SDS
- Limitations are deferred to supplementary material, which is not included in the provided text
- Diversity is not presented as a significant advancement
Experimental setup
Datasets: DreamFusion gallery prompts
Benchmarks: DreamFusion gallery prompts (83 prompts)
Metrics: CLIP Score GPTEval3D ImageReward running time VRAM
Baselines: SDS CSD VSD ISM SDI Consistent3D DreamerXL
Builds on: Score Distillation Sampling Variational Score Distillation Schrödinger Bridge LoRA threestudio
Closest papers by meaning
- Consistent3D: Towards Consistent High-Fidelity Text-to-3D Generation with Deterministic Sampling Prior (CVPR 2024 · similarity 0.91)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (ICLR 2024 · similarity 0.90)
- HIFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance (ICLR 2024 · similarity 0.90)
- Progressive Rendering Distillation: Adapting Stable Diffusion for Instant Text-to-Mesh Generation without 3D Data (CVPR 2025 · similarity 0.89)
- DreamPropeller: Supercharge Text-to-3D Generation with Parallel Sampling (CVPR 2024 · similarity 0.89)
- Let 2D Diffusion Model Know 3D-Consistency for Robust Text-to-3D Generation (ICLR 2024 · similarity 0.89)
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.