Print-Readiness of AI 3D Models: Beyond the Slicer Pass Rate

What is published

The only quantified pass rate we know of comes from Meshy, a generator vendor. Its print-readiness post (posted May 22, 2026) describes 75 generated models, 15 per tool across five image-to-3D tools and, it says, ten reference-image categories (it names only five), checked for mesh integrity in Materialise Magics, then sliced in Bambu Studio with PrusaSlicer cross-validation on a subset. The failure rule is quoted from the post: a model failed “if any repair dialog appeared, regardless of whether the repair was theoretically resolvable”. It says “Meshy achieved 97% slicer pass rate” and that “Other tools in our test set ranged from 63% to 89% on the same metric”, without naming which tools produced which figure. Its format-support table lists Meshy, Hitem3D, Tripo, Rodin and CSM, but as a feature comparison, not as test results. A separate Meshy comparison post (posted March 5, 2026, updated October 27, 2026) is the one that names the five tools as the 75-model test set, says nothing about 15 per tool or the 63 to 89 percent range, and gives no per-tool pass rates; its later sections refer to Hi3D and Hunyuan3D rather than CSM and say CSM’s site has been offline since January 2026, which the post does not reconcile. All of this is vendor-reported.

Reasons to retest rather than rely on it:

• The authors are the vendor with the best score, by their own description “independent testing by the Meshy team”.
• Competitors appear only in a format table, with no per-tool pass rates, no raw data and no model files.
• The 97 percent figure does not map cleanly onto 15 models per tool (14 of 15 is about 93 percent, 15 of 15 is 100), and the post does not explain how it was derived or say whether all 15 per tool were figurines (one line of the same post ties the 97 percent to “character/figurine models”). Two other Meshy posts (the comparison post and a repair post) scope 97 percent to “75 character and figurine models” in Bambu Studio, which is a different sample description.
• It reports slicer outcomes, not physical prints or wall-thickness results.

None of this means the number is wrong. It means nobody outside the vendor can check it, which is why this page gives you a way to check your own files.

Define pass in three levels

A model can pass the first level and fail the third. Report all three.

1. Level 1: slicer opens and slices clean. No repair dialog, no non-manifold warning, no empty layers. This is Meshy’s definition and the easiest to score.
2. Level 2: geometry fits the printer. The model meets the minimum feature sizes of the intended process (below).
3. Level 3: it prints. A physical print finishes and the result is recognizable at the intended scale.

Geometry that is closed but thin, floating or overhanging fails levels 2 and 3 without ever raising a slicer warning.

Thresholds with an official source

Use these only as the reference points they are. If a figure is not here, no official source we read gives one.

• FDM perimeters. Prusa’s knowledge base (undated) says that at a 0.4 mm nozzle one perimeter is about 0.45 mm wide, two about 0.9 mm, three about 1.35 mm and four about 1.8 mm, and that walls thinner than one perimeter are not printable. Check wall thickness against your own nozzle and line width.
• FDM overhangs. The same article says overhangs of about 45 to 60 degrees print cleanly, and up to 75 degrees on MK4S and CORE One class printers. Other printers may differ.
• Manifold requirement. The 3MF core specification requires model meshes to be manifold with consistent orientation and outward normals, which makes each edge shared by exactly two triangles.
• Resin. Formlabs’ Form 3 specification for Clear Resin at 100 microns lists a minimum supported wall of 0.2 mm, minimum unsupported wall of 0.2 mm, minimum hole diameter 0.5 mm, minimum clearance 0.5 mm and a 5.0 mm maximum unsupported overhang length. Formlabs says figures vary by printer and resin; use your printer’s own sheet. See resin miniatures.
• Scale. STL files carry no unit information, so a model can be 100 times too big or small. The 3MF core specification’s unit attribute defaults to millimeters.

Protocol

1. Fix the input set. Use a prompt or image set that is the same for every generator; the site’s standard prompt set is one option. Include at least 30 models per generator across categories you will actually print (figurines, props, functional-looking parts). Record the generator, model version, plan and date. Thirty is a starting size, not a statistical guarantee; report counts, not just percentages.
2. Export once, as the generator recommends for printing. Note format (STL, 3MF), and scale as imported. Do not repair before measuring.
3. Raw mesh checks. For each file record: manifold or not, number of open boundaries, number of separate shells, bounding-box dimensions in mm, and minimum wall thickness at the thinnest feature. Blender’s 3D Print Toolbox extension lists Solid, Thickness, Distorted and Overhang checks in older manual text, but the panel names in the current extension are unconfirmed; verify in your version. Checking raw AI 3D output covers triangle and watertightness counts.
4. Slice in two slicers. Bambu Studio and PrusaSlicer, or OrcaSlicer, at the intended nozzle and layer height. Record the slicer version, any warning text, whether a repair dialog appeared, and whether automatic repair was applied. Do not count a model as passing if it only passed after the slicer silently repaired it; record that as “repaired”.
5. Check the preview. Look through layers for missing wall segments, floating islands and unsupported overhangs beyond the thresholds above.
6. Print a subset. At least five per generator, chosen before slicing so you do not cherry-pick: Levels 2 and 3. Record outcome (complete, partial, failed), print time, supports needed, and what failed.
7. Repeat after repair. For failures, run the repair route in fixing AI models for the slicer and record whether the model then reaches Level 3. Repair success is a separate metric from raw quality.

Record sheet

Report pass rates as counts with the slicer and version, for example “k of n at Level 1 in Slicer X version Y”, and publish the failing files if licensing allows. That is the part the vendor test lacks.

Using the result

A generator with a high Level 1 rate and a low Level 3 rate needs wall-thickening or scale controls, not repair. One with the reverse pattern needs a repair step. Either way the print-readiness checklist turns this into a per-model check. For generator selection on documentation alone, see the printing generator picks. For how we handle unmeasured claims, see methodology. Back to 3D printing.