Retopology settings and reports

The two halves of a quad remesh that a caller holds: RemeshSettings is what you ask for, and RemeshReport is everything that was measured about what came back. Between them sit ConditioningSettings and ConditioningReport, which are the same pair for stage one — the part that turns broken input into something the rest of the pipeline can read at all.

RemeshStage names the seven stages, RemeshStageTiming says how long one took and how much it handled, Singularity is one irregular vertex, and RemeshGuide is a curve the edge flow should follow.

Turning a triangle soup into an all-quad mesh with clean edge loops — a generated blob, a boolean result from the blockout tools, a scanned part — inside the content build, with no external binary and no round trip through a modelling package.

You want the report as much as the mesh. A remesh has no single right answer, and the difference between one that is usable and one that is not shows up as numbers: how far the output strays from the source, how many irregular vertices there are and where they landed, whether the worst quad is inverted. ⚠ A remesher that cannot tell you it went wrong will be trusted until it embarrasses somebody, which is the whole argument for running this unattended.

You do not want it for a mesh whose topology is already the point. Retopology replaces topology by definition; a mesh that only needs texture coordinates wants the unwrapper instead.

Two fields decide the size and only one of them may be set: TargetQuads is a budget, and TargetEdgeLength is the same statement in world units. One implies the other through the surface area, so giving both is an error rather than a preference.

Adaptivity runs from uniform squares at zero to curvature-driven at one. ⚠ The curvature term is weighted by anisotropy — the difference between the two principal curvatures — and not by curvature itself. On a sphere the two are equal, the weight is zero, and the field is free to be smooth, which is the correct answer. Weighting by magnitude instead is the classic failure: it chases noise on a sphere and produces topology that looks agitated for no reason. Measured on this implementation, mean anisotropy is 0.146 on a sphere against 8.26 on a cylinder, where the corresponding magnitudes are 3.55 and 6.66 — so magnitude says a sphere is half as curved as a cylinder, and anisotropy says a sphere has no direction.

⚠ NonQuadCount being zero is not a nicety, it is what makes the output editable. The modelling verbs are built on the assumption that a loop, a ring and a loop cut are statements about four-sided faces. A result with triangles and pentagons in it has no rings to cut, and the whole vocabulary of MeshOperations stops working on it.

Generated meshes arrive with staircase noise at the voxel frequency, self-intersections, floating debris and non-manifold edges. ConditioningSettings says what may be done about it and ConditioningReport says what was. A caller can refuse to continue on the report.

⚠ Every tolerance here is relative to the bounding box and never absolute. A fixed epsilon is a claim about how big a model is, and this repository has been caught by that more than once — an absolute degeneracy test once declared sixty-four real triangles of a capsule's pole degenerate, and a model built at a tenth of the scale would have lost all of them.

⚠ Shrinkwrap is off by default and reading ConditioningReport.Shrinkwrapped as true is worth treating as a warning. It destroys thin features. It exists for input so broken that nothing else will run.

FillHoles is off by default too, for a different reason: a hole in the input is very often a hole in the subject, and closing one silently is worse than leaving it.

A Singularity is a vertex whose one-ring does not close into four quarter turns — a valence-3 or valence-5 point in the output. They cannot be eliminated: the sum of their indices is fixed by the surface's Euler characteristic, so a sphere must carry eight quarter-turns' worth however good the field is. What can be decided is where they go, and the right somewhere is where the surface genuinely is not developable — a fingertip, a box corner, a sphere's pole, which is where an artist puts them.

⚠ "Zero singularities on features" means zero on the interiors of the feature chains, not off every feature vertex. On a cube every vertex is a feature corner and every edge a feature edge, and eight quarter-turns have to exist somewhere — so the literal reading makes the simplest hard-surface shape unsatisfiable. A corner is exactly where one belongs.

Retopologising a generated mesh and refusing the result if conditioning had to work too hard:

Asking for uniform squares rather than adaptive density, which is what a hard-surface part usually wants:

⚠ KeepGroups only reads a group boundary somebody assigned. It is checked against EditMesh.GroupSource, so a shape out of MeshShapes, a boolean of two of them and a mesh whose material a reader carried across all keep their crease — and a mesh welded straight out of EditMesh.FromTriangles, whose groups are the coplanarity guess, does not. On a faceted surface that guess is close to one group per triangle, which would make every edge of the mesh a hard feature.

Guides are an asset rather than a paint session, so one authored against an earlier version of a model still applies:

• Attribute transfer — stage seven, and what stops a remesh being a downgrade.
• Map baking — the normal and displacement maps the atlas is filled with.
• UV packing — where the atlas comes from, and the packer this shares.
• Edit meshes — the kernel a remesh hands quads back to.
• Mesh booleans — the hard-surface case, and what a blockout retopologises.
• Triangle tree — the acceleration structure the attribute transfer asks.