Fast 3D model generation with AI is now possible with LGM.
3D content generation with an AI solution has seen some significant progress in recent months. A few weeks ago, I wrote an article about Genie from Luma Labs, an impressive AI solution for creating 3D models from text descriptions.
However, one drawback with Genie is that the generation process can take around 10 minutes to complete.
Now, a new solution named LGM promises to slash that time dramatically, generating 3D models in just 5 seconds.
How true is that? Let’s find out.
LGM (Large Multi-View Gaussian Model) represents a novel framework that produces high-resolution 3D models from simple text descriptions or even single images.
It uses an asymmetric U-Net backbone along with multi-view diffusion models. This tech combo enables near-instant 3D creation, addressing the resolution and fidelity issues found in earlier models.
If you want to know more about how LGM works, check out this white paper.
Head over to LGM’s HuggingFace page and start typing the description of the 3D object you want to generate. For instance, I wanted to generate a 3D model of a teddy bear.
Prompt: teddy bear
Here’s a preview of the generated 3D model:
This 3D model looked fantastic! All the angles seem to be correct, and despite being viewed from different angles, the model stays consistent—no weird distortions.
Now let’s try a more complex model and see how well the AI can generate it.
Prompt: Ducati Hypermotard 939
Unfortunately, while the separate preview images appeared promising, the complete 3D model wasn’t as successful.
Aside from the text-to-3D capability, LGM also supports 3D object generation from an input image.
The generated 3D model maintained its shape and structure impressively well, another sign of the AI’s understanding of form.
Here are more examples from LGM’s project page:
Imagine how this AI solution could transform several industries.
If you think this is a novel technology, it is not. There have been a few AI-powered 3D model generators out there. A few that I have already covered here on Medium are the following:
This competition is a great thing! It fuels innovation and drives the researchers to make their tools even better, faster, and more accessible.
Overall, I am happy to see yet another AI tool that can generate 3D objects with nothing but a text description or an image. This is my first time seeing a solution that can take an image as an input and generate a cohesive and decent-quality 3D model as an output.
While it’s crucial to remain mindful of the potential challenges this technology poses, its sheer creative potential is undeniable. LGM is truly a leap forward, and I’m excited to see how the developers and researchers can improve their solution to actually generate a downloadable 3D model similar to Luma Lab’s solution.
Software engineer, writer, solopreneur