Retrieval-augmented Text-to-3D Generation

Junyoung Seo, Susung Hong, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Doyup Lee, Seungryong Kim

Keywords: diffusion models, NeRF, 3d synthesis

Abstract: Text-to-3D generation using using neural networks has been confronted with a fundamental difficulty regarding the scale and quality of 3D data. Score distillation sampling based on 2D diffusion models addresses this issue effectively; however, it also introduces 3D inconsistencies that plague generated 3D scenes due to a lack of robust 3D prior knowledge and awareness. In this study, we propose a novel framework for retrieval-augmented text-to-3D generation that is capable of generating superior-quality 3D objects with decent geometry. After we employ a particle-based variational inference framework, we augment the conventional target distribution in SDS-based techniques with an empirical distribution of retrieved 3D assets. Furthermore, based on the retrieved 3D assets, we propose the two effective methods: a lightweight adaptation of a 2D prior model for reducing its inherent bias toward certain camera viewpoints, and delta distillation to regularize artifacts of generated 3D contents. Our experimental results show that our method not only exhibits state-of-the-art quality in text-to-3D generation but also significantly enhances the geometry compared to the baseline.

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/2024/AuthorGuide.

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: 2402