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

LGM (Large Multi-View Gaussian Model) is an open-source 3D generator from the 3DTopia team. The tool tackles the text-to-3D task, i.e., it creates volumetric scenes from a text description. The authors present the development as a model for creating highly detailed 3D content, recognized at the ECCV 2024 conference.

What it does

• Generates 3D models from a text description (text-to-3D).
• According to the repository tags, image-to-3D mode is also supported.
• Uses a Gaussian Splatting-based representation, which, as described by the author, aims for high-resolution output.

What you need to run it

• Platform: local execution, code written in Python using the PyTorch library.
• Memory: the largest model weights file is 1.5 GB, the full set of weights is 4.6 GB. The authors do not specify system RAM or VRAM requirements.
• License: both code and weights are distributed under the MIT license.
• Repository: github.com/3DTopia/LGM, the model is available on huggingface.co/ashawkey/LGM. The project's demo page is me.kiui.moe/lgm.

Who it suits

The tool will be useful for researchers and developers who want to integrate 3D generation into their pipelines or study a modern approach to the text-to-3D task. The open MIT license allows using the code and weights in commercial projects. The image-to-3D mode expands the use cases.

It will not suit those looking for a ready-made application with a one-click interface or a cloud service without the need for local setup. Running it requires a Python environment and PyTorch, as well as sufficient VRAM, which the authors do not specify — you will have to determine it experimentally.

LGM offers an open foundation for experimenting with 3D content generation. The code and weights are available for download, with the latest repository update dated August 2024.