Fill types

Every patch is filled by one of four routes. Auto picks; you can override per patch from the tool header while patches are selected.

Strategy	Produces	Good for
Auto	whichever of the three below the classifier picks; clears an override	the default — start here
Smart fill	mixed quad/tri interior with local, span-driven density	curved and irregular faces, anything with features or cut-outs
Quad	structured Coons quad grid	clean four-sided curved patches
Loft	ruled quad loft strip	bands and sweeps between two clear sides

Auto can change its mind

Auto judges by the result, so tweaking a boundary curve can make a patch come back as a different strategy than before. That is correct behaviour, not a regression — the input changed. Press a strategy to pin it; press Auto to hand the decision back.

Two things hold for all of them:

• The boundary is CAD truth. It is refit onto the surface and kept rigid, at the span count its curves were given. A fill never argues with the boundary, which is why seams weld to zero.
• Density comes from the boundary. Change a curve's spans and the interior follows — on every strategy.

Smart fill

The universal route. Boundary untouchable, interior free: the patch is unwrapped flat, interior vertices are placed at the density the boundary spans dictate, everything is connected, pushed back onto the surface, and triangle pairs that form honest quads are merged.

Five steps, so you know what the knobs actually move:

1. Unwrap — LSCM. Least-squares conformal mapping flattens the patch, preserving angles locally (the same maths as Blender's UV unwrap). The local stretch factor is remembered, so density is measured in true 3D distance rather than distorted flat distance.

2. Interior vertices — Steiner points. Vertices that are not on the boundary. Where they go is the whole question, and it is where the add-on's own work sits:

• a graded lattice whose row and column spacing follows local boundary density — dense near a dense edge, sparse near a sparse one. Neighbouring patches therefore agree on density by themselves, because their spans already agree at the seam;
• span snapping — a column prefers to land on the line between two boundary vertices facing each other across the patch, so the triangulation lays a through-line from vertex to vertex and the rows meet the boundary cleanly;
• a degeneration lever — the D knob divides the target step. At the coarse end the interior collapses to a few stations, then one cross loop — and at the axis's explicit loft terminal the fill hands over to real wall-to-wall rungs (see The loft terminal below).

3. Connect — constrained Delaunay triangulation. Delaunay's empty-circle rule gives the least-sharp triangles possible; constrained means the boundary edges are honoured vertex for vertex. Blender's own implementation does this step.

4. Back onto the surface. New vertices are projected back through the same unwrap. Boundary vertices are never recomputed — they are taken verbatim.

5. Merge into quads. Adjacent triangle pairs that form an honest quad (near planar, angles near right) are merged. On a regular lattice that gives a quad middle, while triangles stay where they belong: at density transitions and around features.

Every switch in the N-panel Fill options box is a lever on these steps (recombine, local density, rails, rungs …); each is listed with its default in Settings. Manual rails (Shift+W in PATCH) add a line a whole station row must seat on; they never re-aim the lattice.

The loft terminal (Rungs)

With the Rungs toggle on (Fill options, off by default), the D knob's coarse end gets a real destination. Drag one density axis decisively past its floor and it snaps to loft — and the contract is exactly the gesture: whichever handle you pull is the direction that lofts.

At the loft terminal that axis carries no interior structure at all — no stations, no median line, nothing for G to slide. Instead the two walls bounding the patch in that direction are bridged rung for rung: one clean edge per vertex of the sparser wall, straight to its counterpart on the denser one, which absorbs its extra vertices as honest triangles. Density and pairing come entirely from the boundary spans — the fill stays subordinate to the curves, so seams still weld.

The law stands down where it would have to guess: end caps and corners keep free triangles, a notch mid-wall fences off its own neighbourhood, and an ambiguous rim (L-shaped, branching) gets the plain fill unchanged.

Structured Quad and Loft

When a patch really is a clean four-sided curved piece, a structured grid beats a graded fill: perfectly regular rows, no triangles. Quad builds a Coons grid from the four sides; Loft rules a strip between two of them.

The trade-off is that their interior is a direct consequence of the boundary, so they have no interior knobs — no G/R/S/D, and the flow guide has nothing to orient (it only sets a lattice frame, so it does nothing here, nor on a flat patch). You steer them by their curves: spans, contour, side roles and core span (see CURVE). Auto escalates to them only when the result is actually clean; a dirty structured attempt falls back to smart fill.

Triangles are not a failure

Flat faces, features and cut-outs are filled with triangles on purpose. Quads earn their place on curved surfaces where they carry a silhouette and deform predictably; forcing them onto a flat plate with a bolt hole just buys you a worse plate. A mixed result with variable spans across a concave boundary is the design, not a shortfall against an all-quad tool.