Module purpose

It starts as a guessing game the moment you open the material slots. You download a 3D model, open it in Blender or import it into Unity, and the slots are gray. The texture folder is a mess: body.png, tex_01.jpg, 57A47F64.dds, mat_30_basecolor, plus embedded maps, external links, dead references, and duplicate packs. Which one is Base Color? Which is Normal? Should Opacity be connected? Engineers try one by one, guessing. Guess right and it looks fine; guess wrong and metal becomes plastic, transparency becomes solid, and it's back to square one. Assets from the internet are even worse—chaotic naming, multiple maps, nested packages. Every import turns into an unpleasant guessing game.

M1 solves this with a clean pipeline: detect → infer → preserve → process → Gate B recheck → package. The LLM identifies the final material identity; the chain plan is the single source of truth. Gate A verifies source hashes and copies only approved resources. Naming is standardized as {Model}_{FinalMaterial}_{Type}.png, real material connections are rebuilt, Metallic and Roughness remain separate, and a complete external Textures/ closure is produced. The result: textures land in the right slots, materials are complete, naming is consistent and traceable, and the asset is ready to import—no more guessing, far less manual trial-and-error.

Module purpose

A one-meter airplane or a hundred-meter giant? Both are everywhere in 3D assets. Units are a mess, and once imported into an engine there's no reliable ruler—so engineers scale each model by feel, through trial and error. Worse, multi-level scaling: root node 100x, child 1.8, grandchild 0.05. Different engineers tweak scales for different needs, nesting chaos until nobody knows the true size.

M2 lets the LLM reset the model to real-world absolute dimensions. The detection stage converts the bounding box to meters. The LLM identifies the object type and true height; a single scale_factor is derived mechanically. Blender applies one unified matrix to the whole assembly, rebuilding the hierarchy through N-level static baking while preserving parent-child relationships, material slots, node connections, and UVs. Final object-level scale is normalized to (1,1,1). Gates verify bounding box, relative distances, hierarchy, normals, and UV facts.

The result: an airplane has its real wingspan, a character has its real height, and it's correct the moment it enters Unity or Unreal. No more ruler-less manual scaling—and those root/child/grandchild scale bombs are gone for good.

Module purpose

In the same project, texture counts are all over the place: some models come with Albedo, Normal, Metallic, and Roughness; some have only a Base Color; others carry a dozen chaotic, unclear maps. Once imported, metal doesn't look like metal, roughness is a guess, and visual quality is uneven. Artists and TAs end up patching maps, wiring nodes, and tweaking parameters per material—and the project starts from an uneven baseline.

M3 consumes the frozen M2 artifact. Under the approved Stage 2 plan, it completes the missing canonical Normal, Metallic, and Roughness for each final material and binds them to the Principled BSDF. Base Color inherits the normalized result from M1/M2. Gate B verifies readability, naming, color space, and hashes. Gate C re-imports the FBX and confirms real node connections, texture closure, and zero UV changes.

The result: whether the source had one map, four, or a dozen, every final material ends up with four standard PBR maps—Albedo, Normal, Metallic, Roughness. Consistent naming, clear channels, real connections, ready on import. A unified quality baseline that isn't dragged down by the worst asset. Failures aren't disguised; anything needing review is clearly flagged.