Researchers Develop GaussianGPT: An Autoregressive Model for 3D Scene Generation

A team of researchers has unveiled GaussianGPT, a novel autoregressive model designed to generate and edit 3D Gaussian scenes token by token. Unlike traditional 3D modeling approaches, GaussianGPT leverages a transformer architecture inspired by GPT models, enabling it to construct scenes from scratch or enhance existing ones while preserving their stylistic and structural coherence. The model's ability to operate sequentially—token by token—sets it apart in the field of 3D scene synthesis, offering a scalable and flexible solution for both creation and modification. The project's source code is expected to be released on GitHub in the coming weeks, with the research community anticipating its potential applications in text-to-scene generation and 3D editing.

Resources: nicolasvonluetzow.github.io