CGGS is a text-to-3D framework for ego-centric scene generation, where limited view overlap and strong viewpoint bias often cause inconsistent content and distorted geometry. It combines an Ego-centric Generator, obtained by fine-tuning a multi-view latent diffusion model with a consistency-augmented loss, with a Layout Decorator that uses optical flow and point-track correspondence to estimate depth and create dense point-cloud layouts. A Geometric Refiner then improves 3D Gaussian reconstruction using an entropy-based Mutual Information Depth Loss and hierarchical optimization. The abstract reports better coherence and geometric accuracy than prior methods, but provides no quantitative results.
No heat snapshots are available in the last 24 hours.