Trellis is a large-scale 3D asset generation model that converts text or image prompts into high-quality 3D assets in multiple formats including Radiance Fields, 3D Gaussians, and meshes.
The model addresses the challenge of scalable, versatile 3D generation by introducing a unified Structured Latent representation that enables decoding to different output formats without retraining. This representation is processed by Rectified Flow Transformers, which serve as the backbone architecture. The approach is grounded in large-scale pre-training on a diverse dataset of 3D objects, allowing the model to learn generalizable patterns for 3D asset creation from various input modalities.
Developers should consider Trellis if they need flexible 3D generation that supports multiple output formats and local editing capabilities. The tool suits projects requiring high-quality 3D assets from text or image inputs, particularly those where output format flexibility matters. The README notes that text-to-3D generation is less creative and detailed due to data limitations, and recommends using text-to-image models as an intermediate step before applying the image-conditioned variant for better results.
The project maintains active development with recent releases of training code and pre-trained models. The team has published both the TRELLIS-text and TRELLIS-image model variants alongside example scripts and a Gradio interface for interactive use. A large-scale dataset with associated data preparation toolkits has been released to support reproducibility and further research. Multi-image conditioning support was added through a tuning-free algorithm, expanding the model's input flexibility without requiring specialized model training.
