Arbitrary-Scale 3D Gaussian Super-Resolution


AAAI 2026


Huimin Zeng      Yue Bai      Yun Fu     
Northeastern University    

TL;DR: We make the first attempt to achieve 3D super-resolution of arbitrary scale factors with a single 3DGS model, providing a unified and efficient solution for flexible, high-resolution novel view synthesis (85 FPS at 1080p).

Abstract

Existing 3D Gaussian Splatting (3DGS) super-resolution methods typically perform high-resolution (HR) rendering of fixed scale factors, making them impractical for resource-limited scenarios. Directly rendering arbitrary-scale HR views with vanilla 3DGS introduces aliasing artifacts due to the lack of scale-aware rendering ability, while adding a post-processing upsampler for 3DGS complicates the framework and reduces rendering efficiency. To tackle these issues, we build an integrated framework that incorporates scale-aware rendering, generative prior-guided optimization, and progressive super-resolving to enable 3D Gaussian super-resolution of arbitrary scale factors with a single 3D model. Notably, our approach supports both integer and non-integer scale rendering to provide more flexibility. Extensive experiments demonstrate the effectiveness of our model in rendering high-quality arbitrary-scale HR views (6.59 dB PSNR gain over 3DGS) with a single model. It preserves structural consistency with LR views and across different scales, while maintaining real-time rendering speed (85 FPS at 1080p).

Drag to Compare

Select methods for left and right videos. Drag the slider to compare the results.

Chair

Microphone

Method

Scale-aware Rendering

Changing output resolution alters image-plane sampling density. A lower sampling rate can fall below the Nyquist frequency and cause aliasing. To align the Gaussian signal with the pixel area, we propose 3D Scale-Aware Smoothing Filter and 2D Scale-Aware Mip Filter to adjust the Gaussian bandwidth and per-pixel integration window, respectively.

Generative Prior-guided Optimization

For Arbi-3DGSR, only input LR views are available. Generative priors from pretrained diffusion models are leveraged to constrain fine details in the rendered HR views.

Progressive Super-Resolving

To remain structurally consistent with LR views, training is divided into multiple stages to progressively accommodate various scale factors.

Visual Results

Blender Dataset (synthetic)

Mip-NeRF 360 Dataset (realistic)

BibTeX

@article{zeng2025arbitrary,
      title={Arbitrary-Scale 3D Gaussian Super-Resolution},
      author={Zeng, Huimin and Bai, Yue and Fu, Yun},
      journal={arXiv preprint arXiv:2508.16467},
      year={2025}
    }