The 3D game asset pipeline, with AI in it

The traditional pipeline

Studios vary, but a 3D game prop typically moves through these stages:

1. Concept. Sketches, paintings, reference boards.
2. Blockout. A rough shape in the engine at the right size, to test gameplay and composition.
3. High-poly model. The detailed version, often sculpted.
4. Retopology. A clean, lighter mesh built over the high-poly. See retopology.
5. UV unwrapping. Flattening the surface so textures can be painted on it. See UV mapping.
6. Baking. Transferring high-poly detail into normal and ambient occlusion maps on the low-poly.
7. Texturing. Painting base colour, roughness, metalness and so on. See PBR materials.
8. LODs. Lower-detail versions for distance. See level of detail.
9. Export and import. FBX or glTF out, engine import settings, materials hooked up.
10. In-engine set-up. Collision, pivots, prefabs or blueprints, lighting checks.

Stages 3–7 are where most of the hours go.

Where AI fits

An AI 3D generator like Vi3W takes a concept (sketch, image or prompt) and produces a textured, retopologised mesh. On the map above, that covers a lot of ground:

• Concept (1): still yours. AI image tools can help explore, but the decisions are the art direction.
• Blockout (2): still yours, and arguably more important, because it sets the size and gameplay purpose of every asset.
• High-poly to texturing (3–7): this is what generation compresses. You get a textured mesh with PBR materials, and Vi3W retopologises it automatically.
• LODs (8): often still a step for you or the engine. Godot can generate LODs on import; Unreal's Nanite changes the picture for static meshes.
• Export and set-up (9–10): still yours, and still where assets succeed or fail.

What changes in practice

• Iteration moves earlier. Because a model costs minutes rather than days, you can try three versions of a prop in the level and keep the best one. Spend that freedom at the concept and image stage, not by generating endlessly at the end.
• Review gates matter more. A traditional pipeline has natural checkpoints (blockout approved, high-poly approved). With generation, add your own: is the image right, is the mesh within budget, does it read in-engine?
• Clean-up becomes a stage. Expect a short pass on many generated assets: scale, pivot, a texture fix. Cleaning up an AI-generated model covers the usual steps.
• Hero assets keep the full pipeline. Characters that deform, and anything seen up close for long, still benefit from hand retopology and careful UVs.

A practical hybrid pipeline for a small team

1. Concept the asset (sketch or reference).
2. Blockout in-engine with a primitive at the right size.
3. Generate from the concept. Check the image stage in Vi3W, then the mesh.
4. Gate: shape right? Triangles within budget? Textures usable?
5. Clean up in Blender if needed.
6. Export FBX or GLB.
7. Set up in-engine: collision, pivot, LOD, prefab.
8. Gate: looks right under game lighting, at gameplay distance.
9. Log how the asset was made (for store disclosure and your own records).

Where the bottlenecks move

With modelling compressed, the slow parts become the ones nobody can generate: art direction, deciding what the level needs, technical set-up in the engine, and consistent style. Those are craft problems, and they're the reason a pipeline exists in the first place. How to make game assets with AI turns this map into a step-by-step routine.

If your pipeline's bottleneck is stages 3–7, try one prop through Vi3W and time it against your usual route.