AESplat: Advancing Pose-Free Feed-Forward 3D Gaussian Splatting via Decoupled Appearance Modeling

2026-09-29

AI Summary

This study addresses the limited rendering quality of existing pose-free feed-forward 3D Gaussian Splatting methods, which stems from their failure to disentangle view-independent and view-dependent appearance. To overcome this limitation, we propose a decoupled appearance modeling framework grounded in spherical harmonics analysis. Specifically, zero-order coefficients are directly derived to represent the base appearance, while an MLP incorporating 3D-aware inductive biases predicts higher-order coefficients to capture view-dependent variations. Evaluated on the RealEstate10K dataset, the proposed method achieves superior performance, improving PSNR by 0.8 dB and 1.1 dB over NAS3R and DepthSplat, respectively. These results demonstrate that explicitly disentangling appearance components effectively enhances novel view synthesis quality in pose-free settings.

Abstract

Pose-free feed-forward 3D Gaussian Splatting (3DGS) has demonstrated remarkable potential for generalized novel view synthesis. However, existing methods typically predict Gaussian appearance attributes represented by spherical harmonics (SH) in the same manner, overlooking the fundamental distinction between view-independent and view-dependent appearance, which results in suboptimal rendering quality. In this paper, we present AESplat, a novel and general framework for pose-free feed-forward 3DGS that introduces an effective decoupled appearance modeling strategy based on an analysis of SH, enabling higher-quality rendering. Specifically, AESplat directly derives the zeroth-order SH coefficient, which represents the base view-independent appearance component, from the input images without training. The higher-order SH coefficients are subsequently predicted by a shallow multilayer perceptron equipped with two efficient 3D-aware inductive biases to model view-dependent appearance variations. Extensive experiments across multiple datasets demonstrate that our method significantly outperforms state-of-the-art approaches, achieving a $0.8$ dB improvement in PSNR over the pose-free method NAS3R and a $1.1$ dB improvement over the pose-required method DepthSplat on the RealEstate10K dataset. Project page: https://aesplat.github.io/.

Problem

• 3D Gaussian Splatting
• pose-free feed-forward
• novel view synthesis
• appearance modeling
• spherical harmonics

Innovation

• 3D Gaussian Splatting
• Decoupled Appearance Modeling
• Pose-Free Feed-Forward
• Spherical Harmonics
• Novel View Synthesis