Retopologize and Unwrap a Generated Mesh

A generated mesh usually looks correct and behaves badly. The silhouette is right, the surface is plausible, and then it will not deform, will not texture cleanly, and will not fit a real-time budget. The reason is topology: how the polygons are arranged, as distinct from what shape they describe.

Reconstruction and generative methods optimise for surface accuracy. They have no reason to produce the edge flow a rigger needs, and they do not.

What is actually wrong with generated topology

Triangle soup. Millions of triangles distributed by surface complexity rather than by importance, with no coherent flow. Nothing downstream can make use of it.

No edge loops. Clean models have loops that follow the forms — around a mouth, an eye, a joint. Deformation follows those loops. Without them, a bend creases and collapses, which is why generated characters cannot simply be rigged.

Uniform density in the wrong places. A flat panel gets the same polygon density as an intricate detail. The flat area wastes budget; the detail may still be under-resolved.

Non-manifold geometry and holes. Edges shared by more than two faces, duplicate vertices, self-intersections, inverted normals. These break booleans, 3D printing, physics, and several export formats.

Floating fragments. Small disconnected islands from reconstruction noise, often invisible until something computes a bounding box.

Whether these matter depends entirely on the destination. A static prop rendered once cares about none of it. A rigged character cares about all of it.

Choose the cheapest cleanup that works

Do nothing when the asset is a static background prop, rendered offline, and never deformed. This is a legitimate and underused answer. If the dense mesh renders and no one will touch it again, retopology is a cost with no return.

Decimate when you need fewer polygons but no structure. It reduces count while approximately preserving shape and leaves the topology just as unstructured. Fine for distant props and LODs, useless for anything that deforms.

Automatic retopology — quad remeshing — produces reasonably even quads and can respect a density guide. This is the right default for hard-surface props and mid-quality assets. It will not produce correct loops around a face or a joint, so it is not a character solution.

Manual retopology is drawing new geometry over the generated surface by hand. It is the only way to get real edge flow, it takes hours, and it is what hero characters and anything with a rig actually need.

The honest test before starting: will this asset deform, will it be seen up close, and will it be reused? Two yes answers justify manual work. Zero justifies leaving it alone.

UVs decide whether texturing works

Once topology is settled, the surface needs a UV layout — a flattening of the 3D surface into a 2D texture space. Generated meshes typically arrive with either no UVs or automatic ones that are technically valid and practically useless.

Four properties matter.

Seams in hidden places. Every unwrap needs cuts. Put them where the eye does not go — under an arm, along a hard edge, behind the object. A seam across a face is a permanent visible line.

Consistent texel density. Every part of the surface should get a similar number of texture pixels per unit of surface area, unless you deliberately give more to an area that deserves it. Mismatched density is why one part of a model looks crisp and another looks blurry, and it is invisible in the UV editor unless you check with a grid pattern.

No overlaps, adequate padding. Overlapping islands means two surfaces share texture pixels, which is occasionally intentional for symmetry and usually a bug. Padding between islands prevents colour bleeding at lower mip levels.

Efficient packing. Wasted UV space is wasted texture resolution.

Check the result with a checker pattern before baking anything. Stretched or differently-sized squares show every problem in the layout in one glance, and thirty seconds here prevents rebaking later.

Practice

Take one generated mesh and prepare it three ways: decimated only, auto-remeshed with UVs, and — if you have the time and the asset deserves it — manually retopologised.

Then attempt one deformation on each: bend it at a joint. The failure modes will be obvious and instructive, and you will never again wonder why a rigger asked for retopology.

Record the time each path took. Those numbers are what let you quote a real schedule the next time someone says a generated model is "basically done."

Common mistakes

Retopologising everything. Static props do not need it. It is expensive work applied to assets that will never benefit.

Auto-remeshing a character and rigging it anyway. The loops are not there; the deformation will crease.

Skipping the checker test. UV problems are cheap to fix before baking and expensive afterwards.

Ignoring scale and orientation. A perfect mesh at the wrong size, facing the wrong way, with its origin in the wrong place is still a broken deliverable.