Scenario Blender Retopology

Blender retopology

Expert retopology is a plan executed with the fewest loops that capture the forms and let the mesh deform: loops close around eyes, mouth and joints, poles sit where nothing moves, and the subdivided surface, not the cage, matches the sculpt. An agent without a mouse can build every piece of this from data, but it reaches expert quality only on the parts it plans explicitly. Automatic remeshing is a finish for static assets and a start for everything else. 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). Toolkit: scripts/bx_retopo.py (import sys; sys.path.append("<skills>/scenario-blender-retopology/scripts"); import bx_retopo as R), all coordinates in the sculpt's local space.

Stance (the expert delta)

• Map first, as data. Kaspar draws six layers on the sculpt before any vertex: edge flow, primary creases, secondary (expression-only) creases, articulation, expression bounds, landmarks (BCON22 [00:07:55], shown in frames 00:07:53 to 00:14:38). The agent writes the same map as coordinates and loop counts (landmarks from renders plus R.surface_point), then builds.
• Loops are the budget, not vertices. Articulation minimum (Kaspar after Brian Tindall): lids, lips and brows as curves with 3 controls and 2 handle loops each, plus 2 corner loops on eyes; upper count = lower count (Dikko, Kaspar, Kenny); 3 loops per joint; a crease is 1 loop on it plus 2 hugging it. Elsewhere "as much as it needs to be" (Kaspar Live #3 [01:56:33]).
• Poles: valence 3 or 5, where nothing moves, never on a crease loop, a shell rim or the face center line (Kaspar Live #7 [00:11:41], Live #3 [00:36:54]; Dikko [00:04:49]). Lampel's reason: pinching is a density jump in the subdivided result, harmless on flat still areas (BCON24 [00:21:18]). [added] An agent's poles come from ring-count reductions (4 per unit) and from cut borders that staircase across an auto base (about half their vertices): budget both.
• Loops close on themselves. A spiral around eyes or lips means rebuild (Dikko); a spiral that terminates in a still area is tolerated (Dikko body [00:09:34], Kaspar nose).
• Judge the subdivided surface. A snapped cage shrinks inside when subdivided; a Shrinkwrap offset fixes convex areas and worsens cavities. Fit the cage so its subdivided surface matches (Duha BCON23 [00:08:58]): R.fit_subdiv. Kaspar's Multires Apply Base does the same job but overshoots (measured below).
• Game low poly = silhouette plus animation loops (SpeedChar): face about 3x body density, triangles wherever nothing bends, twisted quads cut by hand so baker and engine agree, low poly slightly proud of the high; every hard edge is a UV seam (On Mars 3D).
• Auto-remesh fits static, scanned, background or bake-only assets (Lampel [00:09:05], Grant Abbitt, Duha, Jamie Dunbar). For deforming characters it lacks the judgment calls: on the sheep, QuadriFlow at expert density gave no rings around eyes or mouth and tore a thin ear.

Establish first

Workflow

1. Inspect (procedures P1): holes, manifoldness, scale, rotation, symmetry. GATE: all four written down.

2. Map: read the sculptor's face sets first (R.face_set_map, R.face_set_review, P1b: lip line, eye masks, ears, hooves); landmarks with coordinates, loop counts per feature from the Numbers table, crease list (check expression shape keys), per-part density. GATE: map exists before any geometry.

3. Choose a method per part:

   Part	Method	Evidence
   Eyes, mouth, brows (deforming)	carve_rings(corners=, count=(lo, hi)): upper = lower by construction, geodesic rings, the border's count kept when inside the range; pair odd boundaries first (parity_strip)	v2: eyes 24 (12/12), mouth 30 (15/15) on every clean ring; forcing Kaspar's 18/24 cost 40 more poles
   Nasolabial, face frame, neck (map lines)	carve_band along enclosing_contour of face sets (P7): one clean loop, poles pushed onto the base	v2: nasolabial +18 poles; face frame opt-in, +64 poles on the sheep
   Limbs, ears, tail, fingers, horns	limb_profile (geodesic contours, root and joints found) + socket_tube(count=) (P3b, P8); straight parts: tube from sections	v2: all 7 parts incl. the 1 m curled tail, 3 rings per joint
   Skull, torso, smooth masses	QuadriFlow start acceptable, else grid patches	little deformation
   Static prop, scan, bake-only	QuadriFlow whole object (R.quadriflow)	Lampel, Grant Abbitt
   Realistic human head	shrinkwrap a clean generic base mesh, then fit (Kaspar BCON22 [00:20:16])	not scripted here

4. Build islands separately and roughly (Kaspar: "don't focus on tweaking all the points"). Stack via R.setup_cage (Mirror clip > Shrinkwrap Target Normal Project > Subsurf hidden in edit mode). GATE: topology gate below passes per island.

5. Relax and freeze hand-built layouts: R.relax, R.freeze. Not after QuadriFlow (worsened coverage 0.44 % to 0.55 %).

6. Connect by counts: R.boundary_loops, then R.bridge / R.weld_loops on equal counts; reduce the larger side first (Kaspar: wrist 10 vs hand 16 was "six too many"). GATE: no triangles created.

7. Volume: film R.fit_subdiv (gentle defaults); game R.push_outside then R.triangulate_twisted and a Triangulate modifier on export.

8. Review and hand off: gates, two render sheets, handoff hygiene (P13). Report what the agent did not plan.

What an agent reaches (measured on the sheep; tests/code/blender-retopology/sheep_results.json and compare/metrics.json)

Numbers flatter the agent: an aggressive fit scored 0.043 % and looked lumpier than Kaspar's cage. What stays human-level only: articulation counts with handle loops, crease loops placed from expression shapes, pole routing and loop budgets out of the face. Say so in the report.