Taming Gaussian Densification in Score Distillation Sampling for High-Fidelity Text-to-3D Generation
Electronics · 2026-11-15 · 15 (22) · 5310
DOI: 10.3390/electronics15225310 ↗At a glance
Journal article in Electronics, published online on 2026-11-15. Written by Bumsoo Kim and 2 co-authors. 45 references. The full text is freely available (open access, CC BY 4.0). Funded by: Chung-Ang University Graduate Research Scholarship, Korea government, Seoul Metropolitan Government.
Abstract (original language)
In the realm of conditional 3D generation, Score Distillation Sampling has emerged as a promising alternative for generative 3D reconstruction, enabling zero-shot multimodal tasks such as text-to-3D. To address the fundamental limitations of implicit neural representations, such as NeRF and DMTet, follow-up works adopt 3D Gaussian splatting into the SDS framework, leveraging its balanced capability. However, a naive integration without carefully considering the explicit nature of the 3DGS density control strategy leads to fidelity degradation, due to fundamental discrepancy with the target task. To address this issue, we carefully analyze the interaction between the SDS loss landscape and the Gaussian densification strategy. Based on these findings, we propose a Gaussian densification annealing strategy that smoothly regulates the explosive multiplication of Gaussian blobs through progressive knowledge distillation. This idea is inspired by diffusion timestep annealing, which handles coarse-to-fine semantic learning. Moreover, we achieve high-quality 3D generation represented as both 2D images and 3D meshes by implementing an opacity regularization term. Experiments across diverse prompts show improvements in average text–image alignment and qualitative improvements in many examples.
Funding
- Chung-Ang University Graduate Research Scholarship
- Korea government
- Seoul Metropolitan Government
Cite
APA
Kim, B., Kang, Y., & Seo, S. (2026). Taming Gaussian Densification in Score Distillation Sampling for High-Fidelity Text-to-3D Generation. Electronics, 15(22), 5310. https://doi.org/10.3390/electronics15225310
Vancouver
Kim B, Kang Y, Seo S. Taming Gaussian Densification in Score Distillation Sampling for High-Fidelity Text-to-3D Generation. Electronics. 2026;15(22):5310. doi:10.3390/electronics15225310
BibTeX
@article{Kim2026,
author = {Kim, Bumsoo and Kang, Yoonje and Seo, Sanghyun},
title = {Taming Gaussian Densification in Score Distillation Sampling for High-Fidelity Text-to-3D Generation},
journal = {Electronics},
year = {2026},
volume = {15},
number = {22},
pages = {5310},
doi = {10.3390/electronics15225310}
}
EUVIMED