Articles · AI Design
Four in five AI-made 3D models can't be printed as they are
MIT researchers released InstructMesh, a tool for fixing AI-generated 3D models before printing. About 80% of generated models had structural flaws, and novices fixed about 90% with the tool. Here's how the fixing works and what to check before you print.
Translated from the Korean original. Read in Korean
AI that turns a line of text or a single photo into a 3D model is common now. On screen, the results look convincing. Print them on a 3D printer, though, and it's another story. The handle is too thin and snaps, the whistle is sealed inside and makes no sound, and the walls are so thin they collapse mid-print.
On October 1, MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) released a tool aimed at this problem. It's called InstructMesh. It lets you pick only the problem areas of an AI-made 3D model, fix them by describing the fix, and then print.
First, a disclaimer. InstructMesh is still a research tool, so I haven't used it myself. What follows is based on MIT's announcement and the paper, which will be presented at the UIST user interface conference in November.
Right to look at, wrong to use
The researchers' diagnosis fits in one sentence: AI knows what an object should look like, but not how it works.
So first they measured how often that happens. They had the 3D generator TRELLIS recreate popular models from Thingiverse, a 3D printing sharing site.
About 80% were structurally flawed. Four in five wouldn't do their job if printed as is.
Paint the problem, then say the fix
Here's how InstructMesh works. Microsoft's 3D generator TRELLIS makes the model, and the language model GPT-4 understands what the user asks for.
The key is step two. It doesn't rebuild the whole model. You paint only the part with a visible problem, such as the handle or the base. Then you say something like "thicken the handle", or use sliders for small adjustments such as thickness. It mainly does three things.
- Open holes. Open spaces that should be hollow but got sealed, like in a whistle or a bottle.
- Seal holes. Close gaps that shouldn't be there.
- Adjust thickness. Thicken parts too thin to print or that would snap.
The objects the researchers actually printed are fun. A mug wrapped by a dragon whose tail is the handle, glasses with butterfly wings spreading above the lenses, a blue shell-like whistle, an octopus-shaped dispenser, a knee brace that looks like denim. The AI set the shape, and a person did the finishing that made it usable.
When 12 novices tried it, they found and fixed the flaws about 90% of the time, with the results checked by an expert. People who had never done 3D modeling could fix a problem by describing it, as long as they could see where it was.
The same scene, from 2D to 3D
Yesterday I wrote about the week image AI turned toward editing: image models leading with "fix only the chosen spot" instead of "re-roll the whole thing".
InstructMesh goes the same way. Rebuild the whole thing and something else goes wrong besides what you fixed. So you pick the region and touch only that. In 2D or 3D, it's becoming clear that this kind of partial editing is essential to actually using AI output.
The division of labor is the same too. AI produces the shape quickly, and a person judges whether it can actually be used. Will the handle hold when gripped? Will it leak? Where will it break? What product designers have always done matters even more here.
So, for now
InstructMesh isn't a product you can download yet. But if you plan to print from the 3D generator you use now, the three things the researchers fixed make a ready-made checklist.
| What to check | Why | How to look |
|---|---|---|
| Thin parts | They collapse mid-print or snap in use | Thin-wall warnings in your slicer, section view |
| Sealed holes | Things meant to be hollow can't do their job | Cut a section and see if it's hollow |
| Gaps that shouldn't exist | It leaks, or the print breaks off | Mesh check (find holes, flipped faces) |
| How it's used | It has to be gripped, fitted, or stand up | Look at handles, bases and joints separately |
The last row matters most. AI can't imagine a hand gripping the object. Before you hit print, ask once: "Where does the force go when I hold this?" The flaws the novices fixed in the study were ones you could spot by eye. Once you notice them, you can fix them.
3D and mockup tools are collected under Mockups and AI design tools. I wrote separately about bringing AI output into real work in Why thirty AI drafts don't make the work any shorter.
Sources: MIT News, MIT CSAIL, arXiv 2608.28534, 3D Printing.