· OpenAI posts the Point-E paper
Point-E made 3D shapes in minutes by taking the long way round
🤖 AI-assisted, human-curated. How we use AI →
Late on a Friday, UTC time, five OpenAI researchers dropped a paper with a slightly odd plan for making 3D objects from text. Don't go straight to 3D. Make a picture first, then build the shape from the picture.
What the paper said
"Point-E: A System for Generating 3D Point Clouds from Complex Prompts" went up on arXiv on December 16, 2022. The authors are Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin and Mark Chen.
The pipeline has two models. A text-to-image diffusion model makes one synthetic view of the thing you asked for. A second diffusion model then produces a 3D point cloud (a cloud of colored dots in space, roughly the shape of the object) conditioned on that image.
The headline was speed. According to the abstract, Point-E produced 3D models in one to two minutes on a single GPU, where other leading methods typically needed multiple GPU-hours per sample. The authors were upfront that sample quality still fell short of the best methods. They framed it as a trade: one to two orders of magnitude faster, a step down in quality.
I like papers that say what they gave up. This one did it right in the abstract, where nobody can miss it.
Why the detour was smart
Going text to image to 3D sounds like extra work. I'd argue it's less. The image step does the heavy lifting of understanding "a red chair with curved legs," and the 3D step only has to solve one narrower problem: turn this one picture into a shape. Neither model has to do everything.
That pattern shows up everywhere once you notice it. A hard request is often two easier requests stacked on top of each other.
The same move in a chatbot
When people ask a chatbot for something big in one shot (a full report, a whole lesson plan, a 2,000-word story), the model has to make every decision at once. Structure, tone, content, order. It usually picks the safest average for all of them.
Ask for the intermediate thing first and you get a checkpoint. You can fix the outline before a single paragraph exists, which is far cheaper than fixing a finished draft.
R.Solace writes about prompts the way a shop owner talks about orders: what you ask for is what you get.