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

CRM (Convolutional Reconstruction Model) is an open source 3D generator created by the THU-ML team. The tool solves the problem of quickly producing a textured 3D mesh from a single flat image. The authors position it as a solution capable of delivering a result in about 10 seconds.

What it does

• Generates a textured 3D model (mesh) from a single image.
• During the process, it creates intermediate data: a preprocessed source image, six projections from different angles, and so-called CCM (the authors do not expand this abbreviation in the fact sheet).
• Exports the final 3D model to the obj format (as described by the author).

What you need to run it

• Platform and language: The code is written in Python. A local development environment is required to run it.
• Python version: 3.9 (as described by the author).
• License: MIT. The code can be freely used, modified, and distributed.
• Dependencies: The authors recommend installing packages one by one; the full list of dependencies is not provided in the fact sheet.
• Hardware requirements: Not specified in the fact sheet. One can assume a GPU with sufficient VRAM is needed, but the developers do not provide specific figures.
• Repository: github.com/thu-ml/CRM
• Project website: ml.cs.tsinghua.edu.cn/~zhengyi/CRM/

Who it suits

The tool will be useful for developers and researchers experimenting with generative models and single-image reconstruction tasks. The open MIT license allows CRM to be embedded into your own pipelines. The project may also interest those working with rapid 3D content prototyping and willing to run code locally.

It will not suit users looking for a ready-made application with a graphical interface or a cloud service without the need to set up an environment. Also note that the authors do not specify minimum hardware requirements, so you may need to verify compatibility with your existing equipment before running it.

CRM is a research tool published as part of the ECCV 2024 conference. The code is publicly available, with the latest repository changes dated November 2024. The project targets those familiar with the Python ecosystem and ready to work with code directly.