Distilled-3DGS:Distilled 3D Gaussian Splatting
Lintao Xiang, Xinkai Chen, Jianhuang Lai, Guangcong Wang
2025-08-25
Summary
This paper introduces a new method to make 3D models created with a technique called 3D Gaussian Splatting (3DGS) smaller and more efficient without losing quality.
What's the problem?
3D Gaussian Splatting is really good at creating realistic images of 3D scenes from different viewpoints, but it requires a huge number of 'Gaussians' – essentially tiny 3D shapes – to do so. This means the files get very large, taking up a lot of memory and storage space, which can be a problem for using these models on different devices or sharing them.
What's the solution?
The researchers developed a 'knowledge distillation' process. Think of it like a student learning from multiple teachers. They used several versions of the original 3DGS method as 'teachers' – some with added noise or randomness – and combined their knowledge to train a smaller, more efficient 'student' model. They also added a special 'structural similarity loss' to make sure the student model understood the underlying shape of the 3D scene as well as the teachers did.
Why it matters?
This work is important because it allows for high-quality 3D models to be created and used more easily. By significantly reducing the file size and memory requirements, it opens up possibilities for using these models on devices with limited resources, like phones or tablets, and makes it easier to share and distribute them.
Abstract
3D Gaussian Splatting (3DGS) has exhibited remarkable efficacy in novel view synthesis (NVS). However, it suffers from a significant drawback: achieving high-fidelity rendering typically necessitates a large number of 3D Gaussians, resulting in substantial memory consumption and storage requirements. To address this challenge, we propose the first knowledge distillation framework for 3DGS, featuring various teacher models, including vanilla 3DGS, noise-augmented variants, and dropout-regularized versions. The outputs of these teachers are aggregated to guide the optimization of a lightweight student model. To distill the hidden geometric structure, we propose a structural similarity loss to boost the consistency of spatial geometric distributions between the student and teacher model. Through comprehensive quantitative and qualitative evaluations across diverse datasets, the proposed Distilled-3DGS, a simple yet effective framework without bells and whistles, achieves promising rendering results in both rendering quality and storage efficiency compared to state-of-the-art methods. Project page: https://distilled3dgs.github.io . Code: https://github.com/lt-xiang/Distilled-3DGS .