TripoSR and TRELLIS.2 Failed 3D Printing: 9 Out of 20 Models Didn't Make It Through the Slicer

⚡ The Gist in 5 Seconds

• Artyom Kuznetsov ran 20 models from five photo-to-3D generators through Bambu Studio checks — nine didn't pass.
• TripoSR failed 4 out of 4, TRELLIS.2 — also 4 out of 4; Hunyuan3D 3.1 and SAM 3D Objects passed cleanly.
• The test: four images from FLUX.1 [schnell], all generations via API, files used as-is without any touch-ups.

🔍 What Happened

Artyom Kuznetsov, an AI/ML engineer and CTO of a startup building AI technologies for 3D printing, put together a small but nasty test. Four images from FLUX.1 [schnell], each targeting a known weakness of generators: a robot with separate shoulder pads, a thin antenna, and an open claw; a mug with a see-through handle and hollow interior; glasses with wire-thin frames; a dragon with wing membranes and spikes along its back.

There are five generators, and they build geometry in different ways. TripoSR (Stability AI and Tripo, 2024) — an open model using a triplane field, with surfaces assembled via marching cubes. TRELLIS (Microsoft) — a sparse voxel grid and FlexiCubes. TRELLIS.2 — a fresh version with a different shape representation, no field. SAM 3D Objects (Meta) reconstructs individual scene objects from a single image. Hunyuan3D 3.1 (Tencent) — a closed service, its architecture undisclosed. All twenty models were downloaded as-is and run through the same rules Bambu Studio uses to decide whether to show a warning badge. Nine failed.

The numbers are brutal. TripoSR failed 4 out of 4: the mug had 93 separate parts and 612 open edges. TRELLIS.2 also went 4 out of 4: the robot had 3,163 edges where more than two triangles meet, plus 113 parts. TRELLIS stumbled once, on the glasses — a single "pinched" vertex. SAM 3D Objects and Hunyuan3D 3.1 passed cleanly, but SAM's files were full of scattered pieces: the robot — 23, the mug — 13, the dragon — 12. The glasses fared worst: for three of the five generators, the frames tore or fused together. And that's not all — the slicer's first question is whether the model was made in meters. In the file, it's the size of a poppy seed. And the robot's feet float in the air, detached from its legs, even though the preview looks smooth.

🧠 Our Take

Photo-to-3D generators already look ready for prime time: the image is beautiful, the texture is in place, the text on the chest is legible. But printing doesn't check the image — it checks the mesh. Our assessment: the main value of this analysis is that the author measures countable things rather than "does it look right or not" — open edges, vertices with more than two triangles, the number of pieces in a file. Such defects can be spotted in a couple of minutes while the printer bed is still heating up, and they can't be fixed by a pretty render.

What this changes for anyone embedding generation into their software: post-processing stops being an optional bonus. We believe that selling "instant 3D model from a photo" without automatic mesh repair means handing the user a file they'll take to CAD to fix. For the home scenario, the choice is simple: the closed Hunyuan3D 3.1 delivered clean results on all four images, while the open TripoSR and TRELLIS.2 will have to be fixed by hand or with code.

The winners are closed services with tidy geometry, and those who sell not generation but repair. The losers are open models that render beautifully but close surfaces poorly. But hold off on writing them off: the sample is tiny, the images are synthetic, with no background, shadows, or phone-camera perspective.

⚠️ What We Don't Know

It's a single voice — an analysis by one engineer with four images, and the author himself calls this set "easy mode." The check followed slicer rules, not actual printing: "passed" here means "no warning badges," not "will print on the first try." Hunyuan3D 3.1 is closed — its result can't be reproduced and there's no way to look inside. I'd also double-check the generation date from the source — it looks odd.

Prediction: within a quarter, at least one popular slicer will ship with a built-in "model from a generator" check featuring automatic open-edge repair — otherwise the flood of broken STLs from AI services will become a support nightmare.

Disclaimer: one source, a sample of four images, results not officially confirmed.

Source: Habr