UV unwrapping and baking

Expert-level UVs put seams where the texturing method can afford them, set texel density before texture size, and pack with pixel-true padding. Expert bakes measure their projection distances instead of guessing, and are judged in pixels and under a raking light, never eyeballed once. Every stage below has a measurable gate that scripts/bx_uvbake.py computes. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-blender-expert (execution channel, review loop, 5.2 API changes).

Stance (the expert delta)

• Seams follow the texturing method. Kaspar (2021, 2D painting): as few as possible, hidden under hair, clothes and inner limbs. Blender Studio guide (2024, same author): more seams rather than fewer when you paint in 3D or bake, because stretch costs more. Game bakes add a hard rule: every hard edge is a seam, or the normal map cracks (Lampel, Gambrell, On Mars 3D, Polycount).
• Only closed shells and long tubes need extra cuts. Morren: anything closed on itself needs one cut. Gambrell: cut every taped-paper ring once, hidden. But an island with holes (a face with eye holes, a washer) flattens fine. cut_to_disks test-unwraps it and cuts only if it stretches [added; cutting everything folded the sheep's big islands].
• Unwrap method is a measurement, and 5.2 defaults to CONFORMAL. Always pass method=. On Suzanne, Minimum Stretch gives the most even density (spread 1.38), Angle Based the best angles, and Conformal fails by area (5.89). On flat hard-surface islands all three converge. unwrap(method="BEST") picks per island and rejects islands that fold over themselves.
• Density first, texture size second (Huge Menace). Targets: FPS 10.24 px/cm, weapon 20.48, third person 5.12, top-down 2.56. Priority follows the camera: face, hands and front up; soles, inner mouth and back down (SpeedChar, Kaspar). Match texture sets by scaling the denser set down, never the other up (Gambrell).
• Straighten, then equalize, then pack without tilting (Lampel): Follow Active Quads from a truly rectangular quad, pin, re-unwrap the rest, pack with CARDINAL. Padding is a pixel number, res/128 (16 px at 2K), passed as FRACTION margin = gap/(2*res).
• Measure the projection, never guess extrusion, and do not trust one global value. Ryan King's 0.01 then 0.02 was scene-specific. SpeedChar swept six values by hand, and the value that fixed the face broke the armpit ("when we fix something somewhere, some other problems occur", E6QMhGuhyoc [00:26:23]). Rays start on the low, and max_ray_distance counts from the extruded start. The adaptive bake keeps, per texel, the smallest extrusion that hits a front face. On the sheep it cut inverted texels from 7.5% to 1.2% [verified].
• Unique UVs wherever baked lighting differs (SpeedChar: an AO ghost means shared texels). Stacked or mirrored copies move exactly one tile out of 0-1 before the bake (Polycount), because inside 0-1 the last island baked wins.
• A bake is judged in pixels and under light. Blue < 0.5 means rays hit the far side. Exactly flat means rays missed. All-white AO means the distance is too small. Inverted pixels that do not move when extrusion changes point to a geometry problem, not a setting [verified on the sheep]. Multires bakes cast no rays: on the same surface they gave 0.09% inverted texels.

Establish first