Real3D

Model Details:

Model Description:

We use the model architecture provided by TripoSR, which is a Transformer model for 2D-to-3D mapping built on LRM.

We scale it further on in-the-wild image collections by enabling unsupervised self-training and automatric data curation.

• Developed by: Hanwen Jiang
• License: MIT
• Hardware: We train Real3D on 1 node (8GPU) with equivalent batch size of 80 for 5-6 days.

Model Sources:

• Paper: https://arxiv.org/abs/2406.08479
• Project: https://hwjiang1510.github.io/Real3D/
• Code for training and evaluation: https://github.com/hwjiang1510/Real3D

Training Data: Real3D is jointly trained on synthetic data (Objaverse) and in-the-wild image collections. The former prevents training divergence, the latter introduces new knowldege from a broader distribution of real images. We use Objaverse renderings from Zero-1-to-3 and GObjaverse. The in the wild images are from ImageNet, OpenImages, etc.

Misuse, Malicious Use, and Out-of-Scope Use: The model should not be used to intentionally create or disseminate 3D models that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.