CGGS: Consistency-Augmented Geometric Gaussian Splatting for Ego-centric 3D Scene Generation

2026-07-04

AI Summary

This work addresses geometric distortions and semantic inconsistencies in egocentric 3D scene generation, which arise from limited viewpoints and the dominance of individual perspectives. To mitigate these issues, the authors propose Consistency-Enhanced Generative Gaussian Splatting (CGGS), a multi-view diffusion framework that leverages a text-conditioned multi-view latent diffusion model to synthesize semantically aligned 2D images. An initial point cloud layout is constructed by estimating depth through optical flow and point trajectory analysis. The method further incorporates a mutual information–based depth loss and a hierarchical optimization strategy to refine geometric accuracy in 3D Gaussian splatting reconstruction. Experimental results demonstrate that CGGS significantly outperforms existing approaches in visual fidelity, semantic consistency, and cross-view geometric precision.

Abstract

Challenges remain in ego-centric 3D scene generation due to limited view overlap and the dominant influence of individual perspectives on scene interpretation. These factors hinder the creation of viewpoint-consistent and semantically aligned visual content, as well as the construction of accurate geometric structures. In this paper, we propose CGGS, a text-to-3D framework aiming to enhance 3D-content-awareness and address geometric distortions in ego-centric scene generation. Firstly, the Ego-centric Generator is proposed by fine-tuning a Multi-View Latent Diffusion Model with consistency-augmented loss to generate consistent, high-fidelity 2D content aligned with textual descriptions. Then, Layout Decorator leverages optical flow and point-track correspondence to estimate depth, therefore producing dense point clouds as coarse layouts from the ego-centric 2D priors. Building on this initialization, Geometric Refiner is proposed to enhance 3D Gaussian reconstruction via an entropy-based Mutual Information Depth Loss (MID) combined with a hierarchical optimization scheme for improving visual quality and geometric structure. Comprehensive experiments demonstrate that \textcolor{softred}{CGGS} outperforms previous methods in generating coherent and accurate text-driven 3D scenes. Project page: https://cggs-26.github.io/cggs26/.

Problem

• ego-centric 3D scene generation
• view consistency
• geometric distortion
• semantic alignment
• limited view overlap

Innovation

• Consistency-Augmented Generation
• Ego-centric 3D Scene
• Geometric Gaussian Splatting
• Mutual Information Depth Loss
• Multi-View Latent Diffusion