Perceptual Optimization of 3D Gaussian Splatting -Fidelity Rendering

Perceptual Optimization of 3D Gaussian Splatting for High‑Fidelity Rendering

Master internship – Brussel | More than two weeks ago

3D Gaussian Splatting (3

DGS) has recently emerged as an efficient representation for real‑time novel view synthesis, offering a compelling alternative to traditional neural rendering techniques. By representing scenes as collections of Gaussian primitives, 3

DGS enables high‑quality rendering with significantly reduced computational cost. Most existing approaches rely on pixel‑level distortion losses, such as L1 or SSIM, during training. These losses are not well aligned with human visual perception and often lead to overly smooth reconstructions that fail to preserve fine textures and structural details. Recent work has shown that replacing such losses with perceptually motivated alternatives can substantially improve visual quality, with human studies demonstrating a strong preference for perceptually optimized models over standard approaches.

This highlights the importance of designing training objectives that better reflect perceptual quality.

The goal of this thesis is to investigate perceptual optimization strategies for 3D Gaussian Splatting and to analyze their impact on rendering quality. The work will focus on studying different classes of distortion losses, including pixel‑based, perceptual, and distribution‑based losses, and evaluating their effectiveness in preserving fine details and textures. Particular attention will be given to losses inspired by human perception, such as those based on feature statistics or distribution matching.

The thesis will also examine how perceptual optimization affects model characteristics such as the number of Gaussians, rendering efficiency, and compression performance. A secondary objective is to analyze the relationship between perceptual improvements and both objective metrics and human preference, using subjective evaluation as a validation tool.

The research will involve implementing and integrating different loss functions within a 3

DGS training pipeline. Experiments will be conducted on standard datasets to compare reconstruction quality across different optimization strategies. The evaluation will combine objective metrics, such as LPIPS and DISTS, with controlled subjective comparisons to assess perceptual quality. Additional analysis will focus on efficiency‑related factors, including model size and rendering performance. This thesis is expected to provide a systematic analysis of perceptual optimization in 3D Gaussian Splatting and to identify effective loss functions for improving visual quality, offering practical insights for designing perceptually optimized 3D rendering systems.

• Strong programming skills (preferably Python and familiarity with PyTorch or similar frameworks)
• Basic knowledge of computer vision and/or machine learning
• Familiarity with 3D graphics or rendering is a plus, but not strictly required
• Interest in experimental research and evaluation (including perceptual assessment)

• Type of internship:
  Master internship
• Required educational background:
  Computer Science, Electrotechnics/Electrical Engineering
• Supervising scientist(s):
  For further information or for application, please contact Saeed Mahmoudpour ()
• Reference code: 2026-INT-159 – Mention this reference code in your application.

• Resume
• Motivation
• Current study

• Kerbl, Bernhard, et al. "3d gaussian splatting for real-time radiance field rendering." ACM Trans. Graph. 42.4 (2023): 139-1.
• Zhang, Richard, et al. "The unreasonable effectiveness of deep features as a perceptual metric." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.
• Ghildyal, Abhijay, et al. "Non‑Aligned Reference Image Quality Assessment for Novel View Synthesis." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2026.
• Ozyilkan, Ezgi, et al. "Drop‑In Perceptual Optimization for 3D Gaussian Splatting." arXiv preprint arXiv: (2026).