GaussianCube Model Trained on Objaverse

We provide the GaussianCube model trained on Objaverse.

For model usage, please refer to our repo.

The v1.0 is trained under the setting of our paper.

For v1.1 version, we re-filter the data of Objaverse according to aesthetic score. We also include hssd_models and 3D-FUTURE for training, building a training set of around 170k high-quality 3D assets. Moreover, we generate the text captions of each 3D asset using GPT-4o, resulting highly detailed text description. Therefore, our v1.1 model has stronger capability to longer and more detailed input text captions.