The Physna team has published an in-depth breakdown explaining how they accomplished the feat.
Paul Powers, the CEO of Physna, a 3D search and analysis company focused on engineering and design applications in AR/VR, has recently published a lengthy article titled "Creating Equitable 3D Generative AI", offering an in-depth look at the company's text-to-3D model AI.
In this article, Paul explained how a team of three engineers managed to create a generative AI prototype in just a two-week sprint using a dataset of only 8,000 models, discussed the trials and tribulations of the development process, showed some examples of generated 3D models, shared the key takeaways of the experiment, and more. You can check out the full article by clicking this link.
"An unexpected takeaway from these tests is that the toughest part of generative AI in 3D was not generating the model or scene itself, but rather overcoming comparatively simple bugs (like model collision) in the short development time frame," reads the article. "While the decision to only use such a small dataset certainly limited the scope of the prototype, the results proved that generative AI in 3D – both at the object and scene level – can take advantage of the sheer volume of data present in 3D models."