AI texturing is useful when you treat it as a retouch and variation layer, not as a replacement for a PBR library. Generators can invent a plausible brick in seconds. They still ship albedo with baked light, metalness at 0.4, and tiles that crawl under a moving HDRI. The next year of tools will get faster. The production rule will stay: licensed, energy-safe maps first; AI second.
What current tools actually do well
- Upscaling a clean 2K scan to 4K when you need a closer crop.
- Seam hiding and extra tile variations from one photoscanned base.
- Extracting a rough height or normal from a single photo for blocking.
- Dirt, leak, and edge-wear overlays you then multiply onto roughness.
They do poorly at binary metalness, at albedo that survives GI, and at materials that must match a real product. If you need a map primer, use the PBR textures complete guide as the acceptance test: sRGB albedo without lighting, linear roughness, sane normals.
A hybrid workflow that stays shippable
- Pick a CC0 scan from the Locker3d catalog as the legal and visual base. Wood and metal are good tests because generators often fake grain and brushed anisotropy.
- Generate two or three variants. Keep the one that preserves albedo luminance.
- Rebuild metalness as 0/1. Paint dirt as dielectric on top, not as gray metalness.
- Check seams under a rotating HDRI. Setup: HDRI and PBR lighting.
- Choose resolution after the look is locked. Do not upscale to 8K to hide a bad tile. See the resolution guide.
Locker3d files are free CC0, so this hybrid is allowed. Read what you can and cannot do with CC0 before you upload a paid pack to a web generator.
License and training-data traps
CC0 on the input does not control the AI vendor. Some web apps claim a license on uploads. Prefer local tools for client work. Do not generate a “marble that looks like Brand X” for a commercial if you could not photograph that slab. Editorial likeness is still a legal problem when the output is a recognizable product.
Keep a source note: “base: Locker3d Oak_01, roughness variant: local model, 2025-06.” CC0 does not require credit, but a producer will ask where the floor came from.
What is next, practically
Expect better multi-map consistency (one prompt, aligned albedo/roughness/normal) and better tile constraints. Expect engines to ingest those stacks into the same Principled / Default Lit graphs you already use. Do not expect a new shading model that makes roughness optional. Do not throw away photoscans; they remain the calibration target for every generator.
Viewport AI materials that cannot export a standard PNG stack are toys. If you cannot drop the maps into Blender or Unreal, you do not have a texture, you have a preview.
Quality gates before a generated set ships
- Gray-box GI: walls do not go brown from a too-dark floor.
- Chrome sphere: environment is sharp where roughness is 0.
- Metalness histogram: peaks at 0 and 1, not a hump at 0.5.
- One-to-one pixel alignment across maps. A slipped normal is worse than a plain one.
More material practice is on the blog. AI is a force multiplier on a disciplined PBR shelf, not a reason to skip the shelf.
Frequently asked questions
Can AI replace photoscanned PBR textures?
Not as a default. Generators are fast at variation and upscaling. They still miss energy-safe albedo, binary metalness, and seamless tiles under GI. Use AI on top of a real scan when you need quantity, not as the only hero material.
Is it legal to run Locker3d textures through an AI tool?
Locker3d textures are free CC0, so you may modify and create derivatives, including with local or cloud tools. You still cannot claim the original scan is yours. Check the AI vendor’s terms for whether they train on your uploads.
Which maps should I generate versus keep from a scan?
Prefer scanned albedo and normal. Generate extra roughness variation, dirt overlays, or resolution upscales. Rebuild metalness by hand if the generator outputs mid-grays.