Never Mind Goofy Pictures and Homework Help, Can AI Generate True 3D Scenes?

An exercise by Physna proves that it can.

A 3D scene generator, a work in progress, by Physna. (YouTube video by Physna.)

AI has once again captured the public’s attention. This time it’s about ChatGPT, a chatbot that is able to give intelligent, complete and correct answers to essay questions to students to “help” them with their homework. But that does not do ChatGPT justice. It seems to deliver surprisingly good ansBut even the best chatbot, or any AI application for that matter, can help with true accurate 3D part design. While Midjourney has been used by cutting-edge architects such as Stephen Coorlas, its fanciful concepts fall short of being full, accurate 3D models. Luma AI is able to create 3D printable models, but they are not smooth, solid models.

A neural radiance field (NeRF) gets halfway there.

“Think of NeRF as 2.5D,” says Paul Powers, CEO of Physna, who may have found a way to employ AI for full 3D modeling—at least for scenes with furniture.

Powers is familiar to engineering.com readers, having been interviewed for a story on Physna, a smart geometric search application, and as an Engineer Who Matters in 2020.

Powers admits that text-driven generative AI (a category in which chatbots and NeRF are part of) is not his bag but he “wasn’t going to miss out on all the fun.”

His approach to 3D generative design involves quickly assessing existing shapes to see if they fit the text prompt. For example, “Show a room with a table and 4 chairs.” For this, Powers relies on a part’s “DNA,” a descriptor of sorts—a code that captures the essentials of a shape. The part’s DNA allows Physna to search for similar geometry but, as we are about to see, it can also help with creating a scene stocked with real 3D parts.

Powers took on the emerging field of AI for scene creation by doing something different: finding objects that address the prompts rather than digitally draw, paint or reproduce 2D images of them. In order to find them, Powers tapped into a small (8,000 item) library of 3D models. By comparison, Stable Diffusion, a text-to-image application, trained on roughly 600 million images, according to Powers.

This pet project took three Physna developers two weeks to complete.

AI everything except 3D design. (Picture courtesy of Paul Powers, from Medium.)

AI everything except 3D design. (Picture courtesy of Paul Powers, from Medium.)

Powers argument for a 3D shape generator that uses existing shapes is simple. There is a world of 3D objects already created. Why make them all over again? Powers used only the 8,000 3D models available in the Amazon Berkeley Objects (ABO) dataset library. Millions more labeled 3D models can also be found in SketchUp’s 3D Warehouse, SOLIDWORKS 3D ContentCentral and Stratasys’ GrabCAD, to name a few—as well as in countless part libraries maintained by machine and building part manufacturers.wers for any purpose—not just cheating on homework.