Which AI 3D model should you use for game assets?

• Inputs: one stylized prop (wooden barrel), one hard-surface object (sci-fi pistol), one T-pose character (armored knight) — generated once with Image Generation and reused as-is for every engine.
• Settings: image-to-3D, texture on, no polygon target (each engine's default), single image. 3 repeats per input × engine = 54 runs, submitted 6 at a time.
• Measured: success, wall time from submit to result (±5 s, includes queue), faces/vertices from the pipeline's stats file, GLB size, the image-to-3D credit list price at run time — and, from the GLB files themselves, texture count and resolution, UV/normal attributes, vertex and triangle totals, bounding box and pivot offset.
• Not measured: silhouette fidelity, texture quality, topology cleanliness, rig-friendliness, import into Unity, post-processing time — the six criteria above are still yours to judge. The original GLB files are published so you can (the table below says which is which).

• Wall time is measured from the API and includes PicoBerry's queue, so it is an upper bound on engine time and will vary with load.
• PB Slim and PB Slim 2 return low-poly meshes (~4.5–5.3K faces) by default; PB Standard, PB Ultra, Meshy 6 and Hunyuan 3.1 return dense meshes (0.3–2M faces) that you remesh down — pick by what the asset needs next, not by the face count alone.
• n = 9 per engine. Enough to show the shape of each engine's defaults; not enough to rank engines by small differences.

What is inside the files (parsed from the original GLBs, medians of 9)

• Triangle counts parsed from the GLB equal the pipeline's reported faces for all 54 files. Dimensions are in the file's own units (glTF metres): Tripo engines and Hunyuan 3.1 deliver a ~1-unit object, Meshy 6 ~1.9 units, so scale them on import.
• Pivot offset is the distance from the origin to the bounding-box centre. Tripo and Meshy centre the object at the origin; Hunyuan 3.1 places it with the box touching the origin on one axis (offset ≈ half the height).
• Hunyuan 3.1 textures are 4096², the others 2048². Meshy 6 ships four texture images (base colour, metallic-roughness, normal, plus one more), the others three. Nothing here says which looks better — open the files.

Judgment criteria — what this test does and does not fill in

"Not measured" means exactly that — no number or grade is filled in for it. "Inspected" means notes were recorded from the render sheets, not a grade. n = 9 per engine and three inputs: a limited sample, not a general success-rate claim.

Import and mesh quality (Blender 5.1.2, medians of 9 per engine)

• All 54 files import into Blender 5.1.2 without errors (headless, import_scene.gltf). "After weld" merges coincident vertices first — glTF splits vertices along UV and normal seams, so the as-delivered files look non-manifold everywhere; the welded numbers are the real open edges (holes) and loose parts. Both sets are in mesh-quality.json.
• PB Slim and PB Slim 2 deliver low-poly meshes built from many separate parts with open edges (holes after weld in the hundreds, 13–32 parts, ~600–1,200 UV islands). Meshy 6 and Hunyuan 3.1 deliver one closed shell with zero open edges. PB Standard and PB Ultra are one shell with a few open edges. Whether open parts matter depends on what you do next — remeshing closes them (see the post-processing table).
• UV overlap is the share of faces Blender's uv.select_overlap flags; UV area used is the fraction of the 0–1 square covered; texel-density CV is the coefficient of variation of sqrt(UV area / 3D area) per face — 0 would be perfectly even texel density. Islands are only counted below 300k faces.

Post-processing through the PicoBerry API (paid follow-up, Sep 12, 2026)

• Every one of the 54 raw outputs was remeshed to a 20,000-face game budget with pb-remesh (10 credits), then UV-unwrapped with pb-uv (10 credits) on the remeshed result — the two steps a low-poly-plus-clean-UV workflow needs. The 18 character outputs were additionally auto-rigged on the raw output with tripo-rig (15 credits) and meshy-rig (10 credits), rigging only, no clip.
• Times are wall-clock from submit to result including PicoBerry's queue (5 s polling), so they are upper bounds on tool time. Failures are listed as such; the public API does not expose the failure reason, so none is invented here. In this run: pb-remesh failed on 4 of the 9 PB Ultra outputs (the ~1.9M-face inputs) and succeeded on all 45 others; pb-uv failed on 1 of 50 (a PB Slim character); tripo-rig failed on 1 of 18 (a Hunyuan 3.1 character); meshy-rig 18/18. Account balance moved 166,769 → 165,334 credits (1,435) over the run — consistent with the list prices with failed external tasks refunded; the account is shared, so treat this as a reconciliation, not an audit.
• This is one chain at default settings. It does not say a rigged character animates well or that a 20K remesh keeps every detail — open the derived files in your own tool. Post-processing credits per output: remesh 10 + UV 10 (+ rig 15/10 for characters).

Visual inspection — recorded, not graded

For each engine × input, a sheet of the three runs rendered from four fixed cameras (Blender 5.1.2 EEVEE, the file's own PBR materials) was inspected next to the input image and the observations written down: does the shape match, does the colour and material read match, are there visible holes or floating parts, and which axis the model faces. Result: shape matches in all 18 combinations; colour reads close in 13 and deviates in 5 (the barrel's wood tone on the four Tripo engines, the pistol body on Hunyuan 3.1); no holes or floaters seen at 512 px; Tripo engines face ±X, Meshy 6 and Hunyuan 3.1 face −Y in the raw file. Reviewer: PicoBerry product team, AI-assisted inspection of the sheets, Sep 12, 2026 — notes, criteria and limits are in visual-review.json; the sheets are on the CDN so you can check every note. Review sheets (18, one per engine × input): PB Slim stylized-prop, PB Slim hard-surface, PB Slim character · PB Slim 2 stylized-prop, PB Slim 2 hard-surface, PB Slim 2 character · PB Standard stylized-prop, PB Standard hard-surface, PB Standard character · PB Ultra stylized-prop, PB Ultra hard-surface, PB Ultra character · Meshy 6 stylized-prop, Meshy 6 hard-surface, Meshy 6 character · Hunyuan 3.1 stylized-prop, Hunyuan 3.1 hard-surface, Hunyuan 3.1 character

Raw data, the three input images, all 54 thumbnails, the 108 original 3D files (model.glb + preview.glb per run, on a public CDN, SHA-256 in the manifest) and the exact scripts are published so you can re-run or re-check it: summary.json · results.json (every run) · manifest.json — SHA-256, sizes, engine versions, run and asset IDs, GLB download links · mesh-metrics.json — per-file glTF metrics · mesh-quality.json — per-file Blender import, topology and UV metrics · postprocess.json — remesh, UV unwrap and rig results per output · visual-review.json — inspection notes, criteria, limits · README — method · bench.py — the run script (free --summarize mode re-aggregates the published results offline) · mesh-metrics.py — the GLB parser · mesh_quality_blender.py — the Blender metrics and render script · postprocess.py — the API post-processing runner

Review log: Sep 12, 2026 — benchmark run and published (PicoBerry product team, via the public API with a paid account); the guide text above was not changed by the results. Sep 12, 2026 (later) — credits column corrected to image-to-3D list prices (the first version had read the text-to-3D catalog: PB Standard mesh 120, not 60) and reconciled to the account's 21,480-credit balance delta; original GLBs, manifest and glTF metrics published. Sep 12, 2026 (second re-evaluation) — added Blender 5.1.2 import + topology/UV metrics, API remesh → UV unwrap on all 54 outputs and auto-rig on the 18 characters (paid, credits listed), 4-view render sheets and written inspection notes; the criteria table was updated from 'not measured' to what was actually measured or inspected.