Test plan · Generators
AI 3D Texture and PBR Quality: Inspecting Baked Light, Maps and Seams
How to check AI 3D textures for baked lighting, fake roughness and seams under engine lighting, plus what Meshy, TRELLIS.2 and others document about their PBR maps.
A texture that looks good in the generator’s preview can fail the moment it sits in a lit engine scene: shadows painted into the albedo, a roughness map that is flat gray, or seams along UV borders. This page explains how to inspect the maps themselves and how to light them, and summarizes what vendors and the one public texture benchmark we read in full do and do not cover.
Why preview is not the test
Most generators show the model under their own lighting. If light and shadow are baked into the base color, the model looks right there and wrong when you move the sun. Three further failure types matter: no PBR channels at all (a base-color-only file), channels present but meaningless, and visible seams.
What is documented
- Meshy retexture API (docs read 2026-11-08):
enable_pbrdefaults to false and, when true, adds metallic, roughness and normal maps; emission is generated only withmeshy-6and not at 8k (meshy-6-lite,meshy-7andlatestgenerate none).texture_resolutionis 2k, 4k or 8k (4k and 8k needmeshy-6,meshy-7orlatest).remove_lightingdefaults to true but, per the docs, applies only withmeshy-6at 2k or 4k and is ignored withmeshy-7,latestor 8k;enable_original_uvkeeps the model’s existing UVs. So de-lighting is a documented option with a narrow scope; whether it works well is not documented. Multi-view texturing was described in Meshy’s tutorial as base color only. On Free, only untextured Meshy 6 Lite models can be downloaded, so inspecting textured Meshy 7-series output needs a paid plan. - TRELLIS.2 exports
sample.glbdescribed as a PBR-ready GLB, exported OPAQUE by default (README and model card). - 3D AI Studio’s texture generator retextures an uploaded GLB, FBX or OBJ and outputs diffuse, roughness, metallic and normal maps, per its product page.
- Rodin’s Gen-2.5 API docs list a default
materialofPBR(base color, metallic, normal and roughness textures), plusShaded(“Only base color texture with baked lighting”) andHybrid(PBR maps with a base color “that includes baked lighting”), so on those settings baked light is what you asked for. Its resolution table runs from 2K to 8K in the standard texture modes, andextreme-highgives 6K, or 12K with HighPack (vendor documentation).
Vendor lists can disagree with the file you download; always check the maps in the export.
What the glTF spec says each map must contain
Check the maps against the format’s own rules (Khronos glTF 2.0 specification, accessed 2026-11-08), because a map that breaks them will render wrongly in any engine that follows the spec.
| Map | What the spec says | Likely symptom if broken (our reasoning, not from the spec) |
|---|---|---|
| Base color | RGB uses the sRGB transfer function; alpha is linear | Lighting direction visible when viewed alone |
| Metallic-roughness | “Its green channel contains roughness values and its blue channel contains metalness values”; must be linear; red and alpha ignored | Roughness and metal swapped or packed into the wrong channels, so metal looks matte or plastic looks mirrored |
| Normal | Tangent space, linear; green maps to the Y direction | Inverted green gives dents where bumps should be when lit from the side |
| Occlusion | Red channel, linear | Occlusion baked into base color as well, doubling the shadows |
| Emissive | RGB is sRGB; alpha ignored | Glow on parts that should not glow |
Roblox is stricter in its own docs: normal maps must be OpenGL-format tangent space, and roughness, metalness and emissive masks are single-channel 8-bit (create.roblox.com texture specifications), so a map that is fine in glTF may need conversion there. A generator that exports separate roughness and metallic images rather than a packed texture still has to be packed or imported correctly; if the engine import looks wrong, check channel packing before blaming the generator.
What the benchmark covers
Meshy’s blog post comparing Hi3D, Meshy and Tripo (published 7 September 2026, updated 27 October 2026; it states “this is Meshy’s blog, and Meshy is one of the three tools being compared”) describes “meshybench,” a texture benchmark published September 2026 with public code. It scores color, texture style and semantic detail against held-out human-made reference models, plus a mesh-texture agreement metric that checks whether paint follows the geometry. Its limits for this question: it is vendor-run; the metrics as described do not mention roughness or normal maps, baked light or engine lighting; and the post itself notes that its Tripo entry is “labelled 3.5” while Tripo’s public materials name P2.0 and 3.1. It measures something different from what this page proposes. See the tests hub for the other protocols.
Protocol
- For each output in the standard set, open the GLB or the extracted texture files and record which maps exist (albedo, roughness, metallic, normal, emission) and their sizes.
- Baked-light check. View the albedo alone. Directional gradients, dark undersides painted on, or highlights painted on mean lighting is baked. Record yes or no and where.
- Channel check. A roughness or metallic map that is a single flat value, or a normal map that is a flat blue, is present but carries no information. Record flat or varied.
- Engine lighting. Place the asset in a scene and view it under (a) neutral studio light, (b) a low sun from the left, (c) the same sun from the right. Baked shadows stay put when the light moves.
- Seam check. Orbit at close range along UV borders under raking light; record visible seams and stretching, and the location. Fixes: UV and texture fixes.
Fields: maps_present, texture_size, baked_light (yes/no/where), roughness_info (flat/varied), normal_info (flat/varied/absent), seam_visible (none/minor/obvious), stretching (yes/no), white_model (no texture delivered).
Free-tier route and ambiguous results
On Meshy Free only untextured Meshy 6 Lite models can be downloaded, so a free run cannot show textures from that tool. Free options that can: 3D AI Studio’s free signup credits (its getting-started docs say the amount may vary and is typically enough for 3-5 generations), TRELLIS.2 locally if you meet its 24 GB minimum, or any model you already own pushed through the checks above to practice the method. Judge a texture on the file, not the preview.
Rules to set before you start, for borderline readings; they are editorial, not outcomes we have seen.
- Albedo looks flat but the preview looked shaded: the preview was lit; that is a pass for baked light and a note for “shading in preview only”.
- Roughness map is varied but uniformly noisy: varied does not mean correct. Compare against what the material should be (wood against metal) and record “varied, implausible”.
- Seam visible only at extreme close range: record the distance. Pass if it is closer than your camera’s minimum distance.
- Normal map looks inverted under one light: flip the green channel in the engine’s import, retest, and record that the file needed it.
- Opaque export of a glass-like object: TRELLIS.2 exports OPAQUE by default, so transparency is a settings question, not a quality one; record it as “needs alpha”.
Pass or fail
Pass means albedo carries no lighting direction, at least roughness and normal are present and varied where the material should vary, and seams are not visible at your normal camera distance under all three lights. Fail on any baked light that moves wrongly under the lights. A base-color-only asset can still pass for a flat-shaded stylized game, but record it as “channels absent”. For retexturing tools rather than generators, see AI texture generators for 3D; for baking from a high-poly mesh, see baking normal maps from AI high-poly.