Size-aware Compression of 3D Gaussians with Fine-grained Mixed Precision Quantization
Keywords: 3D Gaussian Splatting, Mixed-precision Quantization, Compression
TL;DR: Automatically selecting hyperparameters to compress 3D Gaussians to a target file size while maximizing visual quality
Abstract: In this paper, we propose a method to automatically select hyperparameters to compress 3D Gaussians to a target file size while maximizing visual quality. We iteratively search for a hyperparameter configuration until the file size meets the specified budget. However, existing compression frameworks require completing the entire compression process to determine the compressed file size, which is time-consuming. To accelerate this, we design a tailored size estimator for frameworks that can determine hyperparameters without requiring fine-tuning. Although the finetuning-free frameworks are more predictable, they typically underperform compared to fine-tuning-based approaches, which utilize end-to-end differentiable structures to achieve superior results. To close this performance gap, we propose a mixed-precision quantization strategy that exploits the heterogeneity of attribute channels by compressing each channel with different bit-widths. The resulting combinatorial optimization problem is efficiently solved using 0-1 integer linear programming. Additionally, we partition each attribute channel into blocks of vectors, quantizing each vector based on the optimal bit-width determined in the previous step. The block length is then determined via dynamic programming. Our method identifies hyperparameter settings that meet the target file size within 70 seconds, outperforming state-of-the-art methods in both efficiency and quality. Extensive experiments demonstrate that our approach significantly enhances the performance of fine-tuning-free methods, with its upper-bound performance comparable to that of fine-tuning-required techniques.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 785
Loading