Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

9 of 20 AI-generated 3D models turned out unfit for 3D printing

An AI model can turn a picture into a good-looking 3D figurine in a couple of minutes, but a printer needs a different file: watertight, with no holes and no junk. AI/ML engineer and CTO Artyom Kuznetsov ran four images through five generators and got 20 models. The Bambu Studio slicer rejected 9 of them. The defects include a robot the size of a poppy seed, 93 fragments inside a mug, and feet hanging in the air. That's why the file needs to be checked before printing, not judged by its preview.

AI-processed from Habr AI; edited by Hamidun News
9 of 20 AI-generated 3D models turned out unfit for 3D printing
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1TripoSR and TRELLIS.2 failed all 4 of their models, and the original TRELLIS failed one.
Key metric #2TripoSR's mug has 93 separate parts and 612 open edges.
Key metric #3Hunyuan3D 3.1 produced clean geometry, while SAM 3D had a scatter of extra parts.

You upload a photo of a figurine to a generator, and a few minutes later a neat 3D model is rotating on your screen. You download the file, open a slicer (the program that cuts the model into layers and prepares it for the printer), and there's a robot the size of a poppy seed. That's the most harmless defect.

Artyom Kuznetsov, an AI/ML engineer and CTO of a 3D-printing startup, took four images and five popular generators, got 20 models, and checked them against the rules of the Bambu Studio slicer. Nine of the twenty failed: they had holes, junk inside, the wrong size, or parts hanging in the air. Below is what breaks, why, and how to catch it in a couple of minutes.

How we tested: four images and five generators

The test set was built so that each image targets one known weakness. FLUX.1 [schnell] drew them, since that makes it easier to get a clean background and the needed feature. The robot tests separate parts and a thin antenna. The mug tests a through-handle and an empty interior. The glasses test a frame as thin as wire. The dragon tests thin membranes and sharp spikes.

There are five generators, and they work differently:

  • TripoSR (Stability AI and Tripo) is an open model that builds a field on three planes; the surface is extracted with the marching cubes algorithm.
  • TRELLIS (Microsoft) is an open model based on a sparse voxel grid; the surface is built with FlexiCubes.
  • TRELLIS.2 (Microsoft) is a new version where the shape is defined without a field.
  • SAM 3D Objects (Meta) is an open model that reconstructs individual objects in a scene from a single image.
  • Hunyuan3D 3.1 (Tencent) is a closed service accessed via API; its architecture is not disclosed.

We took the twenty files without any post-processing and checked them against the rules by which Bambu Studio decides whether to show a warning icon. Nine failed. TripoSR failed all four, TRELLIS.2 also failed all four, and the original TRELLIS failed one (the glasses with a "pinched" vertex). SAM 3D Objects and Hunyuan3D 3.1 are formally clean, but SAM has a scatter of separate parts in every file: 23 for the robot, 13 for the mug, 12 for the dragon. The glasses proved the hardest: in three of the five generators, the thin frame either broke or fused together.

Where the defects come from

Most generators build the shape through a field: for each node of a 3D grid, the network decides whether the point is inside or outside the object. Then a triangle surface is stretched over the boundary. The grid works like a sieve: anything thinner than a cell step either falls through or fuses with its neighbors. That's why antennas disappear and close-set legs merge.

The second cause is training. Many models compare renders against a reference from different angles, and such training won't notice a hole that is invisible from every viewpoint.

Here's what we saw in practice:

  1. Scale. STL files carry no units, and all 20 models fit into a cube with a side of roughly 1. The slicer treats this as millimeters. Bambu Studio tries to guess: if the volume is below 8 mm³, it assumes inches (scaling by 25.4), and below 0.008 mm³, meters (by 1000). The robot from TRELLIS.2 ended up one meter tall. No generator got the size right.
  2. Junk inside. TripoSR's mug has 93 separate parts: the mug itself and 92 fragments inside it. The cause is noise in the field: marching cubes wraps every error in a closed bubble. There are also 612 open edges (gaps between triangles), more than half of them micro-gaps invisible to the eye.
  3. Edges shared by four triangles. In a normal mesh, exactly two triangles meet at each edge. The TRELLIS.2 robot has 3,163 such problematic edges, plus 113 separate parts. The original TRELLIS produced a clean mesh of the same robot: the new version captures complex shapes better but damages the geometry.
  4. Floating parts. The robot's feet are separated from the legs by 0.5–0.7 mm and fall 0.15 mm short of the table. The slicer refused to slice the model at all because of an empty layer at the antenna.
  5. Details that exist only in the texture. Hunyuan3D 3.1 produced the cleanest geometry, but the eyes and the chest lettering live in the texture. Without it, only barely visible bumps remain.

How to check in five minutes

The warning icon in Bambu Studio says nothing about size, thin walls, or floating elements, so the order is this:

  • Open the STL without its texture: whatever isn't in the relief won't be in the plastic.
  • Hover over the warning icon and look at the number of open edges.
  • Set the real size along one axis while keeping the proportions.
  • Choose Split → To objects: if the model breaks into dozens of fragments, there is junk inside.
  • Slice and scroll through the layers: an element that starts in mid-air will print into nothing.

For automation, a short Python script with the trimesh library is enough: it computes the dimensions, open edges, edges shared by three or more triangles, and the number of separate parts. On TripoSR's mug it reports 612 open edges and 93 parts; on the TRELLIS.2 robot, 3,163 problematic edges and 113 parts. The script won't catch floating feet, details that exist only in the texture, or the "pinched" vertex.

Minor issues can be fixed in the slicer or in Blender (3D Print Toolbox) and MeshLab, but automatic repair can bite off fingers and antennas along with the junk. Floating feet will have to be moved by hand.

The takeaway for anyone building 3D generation into a product: geometry checks and size normalization should be part of the pipeline, not a manual step before printing. This experiment is small: 20 models from a single generation date, so it's a snapshot, not a ranking. The author promises a follow-up on how the neural network's architecture predicts the pattern of defects and what slicers check.

Story

Key Questions Answered

How we tested: four images and five generators?

The test set was built so that each image targets one known weakness. FLUX.1 [schnell] drew them, since that makes it easier to get a clean background and the needed feature. The robot tests separate parts and a thin antenna. The mug tests a through-handle and an empty interior. The glasses test a frame as thin as wire. The dragon tests thin membranes and sharp spikes.

Where the defects come from?

Most generators build the shape through a field: for each node of a 3D grid, the network decides whether the point is inside or outside the object. Then a triangle surface is stretched over the boundary. The grid works like a sieve: anything thinner than a cell step either falls through or fuses with its neighbors. That's why antennas disappear and close-set legs merge. The second cause is training.

How to check in five minutes?

The warning icon in Bambu Studio says nothing about size, thin walls, or floating elements, so the order is this: Open the STL without its texture: whatever isn't in the relief won't be in the plastic. Hover over the warning icon and look at the number of open edges. Set the real size along one axis while keeping the proportions. Choose Split → To objects: if the model breaks into dozens of fragments, there is junk inside.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…