Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Walking the Schrödinger Bridge: A Direct Trajectory for Text-to-3D Generation
Authors: Ziying Li, Xuequan Lu, Xinkui Zhao, Guanjie Cheng, Shuiguang Deng, Jianwei Yin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate that Tra Ce consistently achieves superior quality and fidelity to state-of-the-art techniques. Our proposed Tra Ce framework, which operationalizes the direct transport path via Schrödinger Bridges, is rigorously evaluated. Extensive experiments demonstrate that this approach yields highfidelity 3D assets with strong adherence to textual descriptions (Figure 4 and Table 1). |
| Researcher Affiliation | Academia | Ziying Li Zhejiang University EMAIL Xuequan Lu University of Western Australia EMAIL Xinkui Zhao Zhejiang University EMAIL Guanjie Cheng Zhejiang University EMAIL Shuiguang Deng Zhejiang University EMAIL Jianwei Yin Zhejiang University EMAIL |
| Pseudocode | No | The paper describes the methodology in Section 4.2 'Trajectory-Centric Distillation' using mathematical equations and textual descriptions, but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | https://github.com/emmaleee789/Tra Ce.git and 'Our code will be released to the community.' from the NeurIPS Paper Checklist, section '5. Open access to data and code'. |
| Open Datasets | Yes | We quantitatively evaluate our Tra Ce against other methods using 83 distinct prompts from Dreamfusion online gallery2 with 120 views per prompt. 2https://dreamfusion3d.github.io/gallery.html |
| Dataset Splits | Yes | We quantitatively evaluate our Tra Ce against other methods using 83 distinct prompts from Dreamfusion online gallery2 with 120 views per prompt. |
| Hardware Specification | Yes | With an average processing time of 14 minutes and an average peak VRAM usage of 18741 MiB, Tra Ce offers high-fidelity generation with a compelling balance of qualitative performance, computational efficiency, and memory footprint. |
| Software Dependencies | No | The paper mentions using a 'Lo RA-adapted model' and refers to 'pre-trained 2D text-to-image diffusion models' but does not specify specific software packages or libraries with version numbers (e.g., PyTorch version, Python version, specific LoRA library version). |
| Experiment Setup | Yes | For a sampled time t [0.02, 0.5]... Throughout the optimization of θ, the time parameter t is progressively decreased from an initial value near 0.5 towards 0.02. This common annealing technique gradually shifts the focus of the Schrödinger Bridge interpolation from broader states towards those more proximate to the estimated ideal target xpred 0 , aiding the progressive refinement of the rendered output g(θ, c). We investigate the impact of the CFG value on our Tra Ce, as illustrated in Figure 6 with two example objects. |