Repairing an AI-Generated Model: Which Fixes Work, and Which Make It Worse

In our test of 16 AI-generated models, one repair made every model watertight and left four of them exactly as hollow as before. So what does fix a generated model? We took the nine most useful models from that test, seven with a defect and two healthy ones as controls, and ran six repairs on each.

What we repaired, and how

The models come from the earlier test, scaled to 80 mm and turned the way a slicer receives them. Each has one known problem:

Model	Generator	Problem
Pumpkin, mug, charger, figurine	TRELLIS.2	Hollow: two surfaces 0.15 to 0.16 mm apart, nothing inside
Mug	Stable Fast 3D	Opening sealed: the mug is solid
Figurine	Hunyuan3D-2.1	A 0.47 mm sheet under the figurine, fused to it
Figurine	Neural4D	8 open edges and a second shell
Mug, pumpkin	Hunyuan3D-2.1	None: the controls

The author of this site works for the company that makes Neural4D. Its model went through the same repairs and the same script as the others, and it appears in the tables but not in the pictures.

The six repairs:

1. Windows repair service, the one PrusaSlicer calls when you choose “Fix through the Netfabb”.
2. Fill holes in the trimesh library, standing in for the hole filling every mesh tool has.
3. Make Manifold from Blender’s 3D Print Toolbox 1.5.0, at its default settings.
4. Voxel remesh in Blender, at 0.5 and 1.0 mm. It rebuilds the surface as the outside of a block of small cubes, which is how a skin can become a solid. On the hollow models we also tried 1.5, 2 and 3 mm.
5. Repair, then remesh: the Windows result remeshed.
6. Solidify in Blender, 1.2 mm inwards at default settings: three lines of a 0.4 mm nozzle added behind every surface.

Every result was measured the way the earlier test measured the originals: closed or not, how solid, the thinnest wall, what PrusaSlicer 2.9.6 made of it, and how far the surface moved. The last one is what tells you a repair made things worse.

Hollow skins: closing is not filling

Here is the TRELLIS.2 mug cut through the middle, as generated, after the Windows repair, and after a 1 mm voxel remesh:

The three one-click repairs left the hollow models as hollow as before. None came out more than 3.9% solid, and some did not even come out closed:

TRELLIS.2 model	As generated	Windows repair	Make Manifold	Fill holes
Pumpkin	1.32 m, not closed	1.31 m, closed	1.32 m, closed	1.32 m, not closed
Mug	0.49 m, not closed	0.49 m, closed	0.49 m, not closed	0.49 m, not closed
Charger	0.98 m, not closed	0.60 m, closed	0.61 m, closed	0.50 m, not closed
Figurine	0.78 m, not closed	0.78 m, closed	2.43 m, not closed	0.78 m, not closed

The metres are PrusaSlicer’s filament estimate. The same objects from the other generators took 5.7 to 23.5 m. Every repaired skin still had walls of 0.15 to 0.16 mm. The charger’s starting figure is not stable either: the same file sliced to 0.50 m in the earlier test. With 401 open edges its inside is undefined, and the slicer did not give the same answer twice.

A voxel remesh is what fills them, when it works. At 1 mm, about 1/80 of the model’s size:

• Mug: one solid shell with a 1.9 mm wall and its handle hole intact: a real mug, at 9.65 m of filament. The healthy Hunyuan3D-2.1 mug takes 11.26 m.
• Charger: solid, 14.36 m.
• Figurine: solid, 8.56 m, but with some detail lost. 99% of its surface stayed within 3.9 mm of where it was, against 0.55 mm when we compared the original with itself.
• Pumpkin: broke into 2,091 pieces, and PrusaSlicer refused it: “There is an object with no extrusions in the first layer.”

The size matters. At 0.5 mm the mug broke into 8,176 pieces instead. The pumpkin never worked: at 0.5, 1, 1.5, 2 and 3 mm it came out in 9,667, 2,091, 718, 338 and 102 pieces, with nothing inside any of them.

Repairing first did not help the remesh. Remeshing the Windows repair’s output gave the same solidity as remeshing the original on every model, and the pumpkin shattered just the same.

Solidify is a partial fallback. It added material to all four skins, 10.7 to 16.3 m of filament including 16.31 m for the pumpkin, the only repair that gave the pumpkin more than a few centimetres. It also kept the shape: 99% of the original surface stayed within 1 mm. But none of the four came out closed, the figurine ended up in 91 pieces, and parts of the charger were still 0.16 mm thick. PrusaSlicer sliced them, but read that as a model it can print, not a model that is right.

Defects no repair fixed

The sealed mug. Stable Fast 3D’s mug is solid where the cavity should be, and it is watertight, so the one-click repairs saw nothing wrong with it. Every repair except Solidify left it 85% solid. Solidify hollowed out the inside to a 1.2 mm shell, but the top stayed sealed, so it would print as a closed box shaped like a mug. A repair cannot find a cavity that is not in the file. Regenerate it.

The sheet. The 0.47 mm sheet under the Hunyuan3D-2.1 figurine is part of the same closed shell, so every one-click repair kept it. A 1 mm remesh split it off as a separate floating piece, which PrusaSlicer flagged as “Floating object part”. Repairing first and then remeshing left a file PrusaSlicer would not slice at all. Delete the sheet by hand in Blender: select a face on it, select linked, delete. Then repair.

Holes: one tool closed them

Neural4D’s figurine had 8 open edges. The Windows repair service closed it, and Repair then remesh at 1 mm also merged its two shells into one. Fill holes, Make Manifold and both remeshes on their own left it not closed. Of all seven models with a defect, only the Windows repair, alone or followed by a 1 mm remesh, produced a closed file every time. Make Manifold closed this figurine’s eight open edges but left two non-manifold ones, so its file was still not closed.

The controls: when repair makes it worse

On the two healthy models, the three one-click repairs and remeshes of up to 1.5 mm stayed within our measurement noise for shape and within 1% for filament. Larger changes were another matter:

• A 3 mm remesh opened 14 extra holes through the healthy mug’s wall, thinned it to 0.6 mm in places, and moved 99% of its surface by up to 2.5 mm. At 2 mm the same mug still came through intact.
• Solidify turned the solid Hunyuan3D-2.1 pumpkin into a 1.2 mm shell: from 93% solid to 10%, and from 23.53 m of filament to 8.16 m. That can be useful, a hollow print is cheaper, but it is a change to the object, not a repair.

So repair a model only when you know what is wrong with it, and check the result.

Which fix for which problem

• Diagnose first. Open the model in our GLB or STL viewer: it reports holes, shells and size. Slice it and compare the filament with what the object should take: a few percent means a skin.
• Holes or loose shells: a one-click repair. The Windows service, as PrusaSlicer's Fix through the Netfabb, produced a closed file for all seven damaged models.
• A hollow skin: voxel remesh at about 1/80 of the model's size, then cut it in half to check the inside is filled. If it shatters, try Solidify, or regenerate.
• An extra sheet or slab: delete it by hand before any other repair.
• A sealed opening: regenerate. No repair adds a cavity that is not in the file.
• A healthy model: leave it alone. Large voxel sizes and Solidify change healthy models too.

The general walkthrough of each repair tool is in how to repair an STL file. Deciding between fixing and starting again is in repair or rebuild. If the model is too big for your printer, repair before you cut: our cutting test found every cutting method carried an open model’s holes into the pieces. The order of repair steps matters for ordinary files too, which we measured in our repair order benchmark.

Method

The nine models are the 80 mm copies from our generated-model baseline, generated between 2026-09-24 and 2026-10-15. Blender 5.2.2 ran the Blender repairs headless. The Windows repair results are the same files as in the baseline. Each result was measured with trimesh, and sliced with PrusaSlicer 2.9.6 at its default profile. Surface movement is the distance from 60,000 points on one surface to the nearest of 60,000 points on the other, in both directions. Two samplings of the same unrepaired model differ by 0.53 to 0.74 mm at the 99th percentile, so read every figure against that. Nothing was physically printed. Every reading and the voxel size sweep are in the data.

The raw output is published as JSON, produced by scripts/measure-gen-repair.py.