AI Research Atlas

TRELLIS (Structured 3D Latents)

Microsoft Research · 2 December 2024

TRELLIS encodes 3D assets in a unified structured latent that decodes to radiance fields, Gaussians or meshes; models up to 2B parameters on 500K objects.

Rectified-flow transformers over a sparse 3D grid with dense vision-foundation features; MIT-licensed weights on Hugging Face. A de-facto open baseline for image-to-3D in 2025.

Date
Monday, 2 December 2024
Lab
Microsoft Research
Kind
model
Access
open weights

arXiv v1 2024-12-02; Hugging Face repo same day (MIT).

Sources

  1. arxiv.org/abs/2412.01506
  2. huggingface.co/microsoft/TRELLIS-image-large

This record was checked against its sources on 6 October 2026. How we check