2Xplat: Decoupling Geometry and Appearance Modeling for Feed-Forward 3D Gaussian Splatting

Hwasik Jeong1* Seungryong Lee2* Gyeongjin Kang2 Seungkwon Yang1
Xiangyu Sun2 Seungtae Nam1 Eunbyung Park1†
1Yonsei University 2Sungkyunkwan University
Paper Code 🤗 Demo
Loading...

Abstract

Pose-free feed-forward 3D Gaussian Splatting (3DGS) has opened a new frontier for rapid 3D modeling, enabling highquality Gaussian representations to be generated from uncalibrated multi-view images in a single forward pass. The dominant approach adopts unified monolithic architectures, often built on geometry-centric 3D foundation models, to jointly estimate camera poses and synthesize 3DGS representations within a single network, entangling geometric reasoning and appearance modeling within a shared representation. In this work, we introduce 2Xplat, a pose-free feed-forward 3DGS framework based on a two-experts design that explicitly separates geometry estimation from Gaussian generation: a dedicated geometry expert first predicts camera poses, which are then passed to an appearance expert that synthesizes 3D Gaussians. Despite its conceptual simplicity, and being largely underexplored in prior works, our two-experts pipeline outperforms prior pose-free feed-forward 3DGS approaches in fewer than 5K training iterations, achieving performance on par with state-of-the-art posed methods. These results challenge the prevailing unified paradigm and suggest the potential advantages of modular design for complex 3D geometric estimation and appearance synthesis tasks.

Note: All 2Xplat results and visualizations on DL3DV use view-dependent opacity unless otherwise stated.

Method Overview

2Xplat follows a two-expert design that explicitly decouples geometry estimation from appearance modeling for pose-free feed-forward 3D Gaussian Splatting. A dedicated geometry expert uses coarse tokens to predict camera poses from uncalibrated input views. Guided by these estimated poses, the appearance expert leverages fine tokens to capture local textures and fine-grained visual details, synthesizing a detailed 3D Gaussian representation. By assigning camera pose estimation and detail-rich Gaussian synthesis to specialized experts, this design improves reconstruction quality and supports more robust feed-forward 3DGS generation.

Method overview figure

Quantitative Comparison

2Xplat consistently achieves higher PSNR than the previous state-of-the-art method across different view settings and backbone architectures while maintaining comparable or significantly faster inference speed, demonstrating its robustness and efficiency. The results shown in the plot below are measured on DL3DV at a resolution of 224×224, with inference speed evaluated on a single NVIDIA RTX 3090 GPU.

Quantitative comparison figure

As shown in the graph below, across all view settings (32, 64, and 128 views) on DL3DV at a resolution of 960x540, our "pose-free" method, 2Xplat, achieves performance that is consistently comparable to "pose-dependent" approaches, with only a small gap to the best-performing baseline. This demonstrates that our model can effectively recover high-quality rendering without relying on explicit camera pose supervision.

Quantitative comparison figure

Qualitative Results

Dataset

RE10K

224x224 6 Context Views
Dataset

DL3DV

224x224 6 / 12 / 24 Context Views
6 Views
12 Views
24 Views
Dataset

DL3DV

960x540 32 Context Views

Qualitative Comparison

Comparison between 2Xplat and YoNoSplat on the DL3DV dataset using 12 context views with a resolution of 224×224. For a fair comparison, 2Xplat does not use view-dependent opacity in this comparison.

BibTeX

@article{jeong2026twoxplat,
  title={2Xplat: Decoupling Geometry and Appearance Modeling for Feed-Forward 3D Gaussian Splatting},
  author={Hwasik Jeong and Seungryong Lee and Gyeongjin Kang and Seungkwon Yang and Xiangyu Sun and Seungtae Nam and Eunbyung Park},
  journal={arXiv preprint arXiv:2603.21064},
  year={2026}
}