Path-Tracing Distillation: Enhancing Stability in Text-to-3D Generation by Mitigating Out-of-Distribution Issues

Zejian Li, Jiahui Zhang, Shengyuan Zhang, Tianrun Chen, Tianyun Zhang, Guang Yang, Changyuan Yang, Zhiyuan Yang, Lingyun Sun

Keywords: Text-to-3D Generation, Score Distillation, Path-Tracing, 3D Model Stability

Abstract: Text-to-3D generation techniques signify a pivotal advancement in creating 3D models from textual descriptions. Contemporary state-of-the-art methods utilize score distillation processes, leveraging 2D priors to generate 3D assets. However, these approaches frequently encounter instability during the initial generation phases, primarily due to an distribution divergence between the pretrained score prediction network and the nascent 3D model. Specifically, raw rendered images of an initial 3D model lie outside the distribution (OOD) of the pretrained score prediction network, which is trained on high-fidelity realistic images. To address this OOD issue, we introduce an innovative Path-Tracing Distillation (PTD) technique that refines the distillation process. Our method sequentially optimizes the 3D model using intermediate score networks that exhibit closer distributional alignment, thereby accelerating the convergence during the early stages of training. This approach not only ensures a more stable increase in CLIP similarity initially but also preserves the visual quality and diversity of the generated models. Experiments demonstrate that PTD significantly enhances both the stability and quality of text-to-3D generation, outperforming existing baselines. PTD can also be generalized to other score distillation methods.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5716