OpenLRM: 3D generator — what it does and what you need to run it

OpenLRM is an open-source 3D generator from the 3DTopia team that tackles the image-to-3D task. The tool produces a volumetric model from a single input image and is positioned by its authors as an open implementation of Large Reconstruction Models.

What it does

As described by the authors, OpenLRM reconstructs 3D geometry from a flat image. The repository states that the code is written in Python using the transformers library. The model is available for download: the largest weights file is 1.7 GB, and the total size of all files is 3.4 GB. The authors mention the openlrm-obj-small-1.1 variant, which was trained on the Objaverse dataset and, according to the documentation, allows exporting the result to the obj format. The developers do not specify other export formats.

What you need to run it

Running locally requires the CUDA platform — this is a direct statement in the documentation: the authors recommend reducing the frame_size parameter in configurations if a CUDA OOM error occurs. The code is distributed under the Apache-2.0 license, and the weights files are under cc-by-nc-4.0. This means commercial use of the weights is restricted. The source code can be found in the repository at github.com/3DTopia/OpenLRM, and the weights at huggingface.co/zxhezexin/openlrm-mix-large-1.1. The latest tagged release is version v1.1.1, with the last code change dated May 2024.

Who it's for

The tool will be useful for researchers and developers in computer vision who want to understand the Large Reconstruction Models architecture or fine-tune the model on their own data. The open code and clear license on the source code facilitate this. Those looking for a ready-made solution for a commercial pipeline should consider the non-commercial license on the weights and the mandatory requirement of a GPU with CUDA support. If you don't have a GPU or it's not from NVIDIA, running OpenLRM locally won't be possible. The authors do not provide a cloud version, so you can only rely on your own hardware. OpenLRM carves out its niche as a research project with a transparent architecture. It doesn't try to be a universal all-purpose tool, but it does let you reproduce and study the approach to reconstructing 3D models from a single image.