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

Input	Why it changes the plan	Default when silent
Purpose	film cage renders at subdivision 2 (Kaspar); game ships triangles and bakes	deforming character: film-style all-quad cage
Budget	film: no vertex budget, minimum loops (Kaspar); game: triangle budget	film stylized head 1,650 to 4,100 faces, body 9,400 to 16,500 (Blender Studio meshes); game character < 10,000 tris
Deformation	face rig, joints, how far the mouth travels (expression bounds)	face + limbs deform
Camera	detail and density follow the closest view (SpeedChar)	close-up face
Style	stylized: creases in topology; realistic: generic base + displacement (Kaspar)	stylized
Symmetry	R.topology_report(sculpt)["sym_local_pct"], local space	Mirror only above ~98 % (the sheep is 21 %, posed)
Separate parts	eyes, teeth, hair clumps, clothing (from the body duplicate)	separate objects

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)

Strategy	Faces	Subdiv-2 mean error	Inside	Eye / mouth rings	Closed-loop edges	Looks
Kaspar's own retopo	2,596	0.24 %	47 %	5,5 / 2	29.9 %	clean, reads unsubdivided
QuadriFlow, area-matched budget	4,372	0.14 %	87 %	0 / 0	1.6 %	faceted, torn ear, uniform density
+ gentle fit_subdiv	4,372	0.045 %	53 %	0 / 0	1.6 %	smooth masses, lumpy hooves
+ Multires Apply Base instead	4,372	0.11 %	15 %	0 / 0	1.6 %	overshoot
Hybrid: eye rings + ear tubes + gentle fit	4,544	0.042 %	52 %	3,3 / 0	5.9 %	first skill version
E1 run with the skill (fresh agent: face sets, warp, 4 eye rings, 5 lip rings, limb tubes)	5,778	0.040 %	51 %	4,4 / 5 (46 to 48 verts)	19.4 %	planned face, rippled lids, lumps at leg joins
v2 tools, final (test_sheep_v2.py: face sets, counts as ranges, corner split, geodesic rings and tubes, nasolabial loop)	4,980	0.028 %	52 %	4,4 clean 24-vert rings, 12/12 / 5 x 30, 15/15	33.9 %	best agent result: poles 5.7 % (Kaspar 5.2 %), none on clean rings, no 6-poles; nose and upper-lip patch irregular (human pass)

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.

Numbers

Item	Value	Source
Render subdivision (film)	2 levels; viewport 1	Kaspar Live #6 [00:21:53]; BCON22 frame 00:26:38
Joint	3 loops over the bending part	Kaspar Live #5 [01:37:48]
Limb start ring	8 verts (legs 8 then 12); Dikko arms 8 min, 10 to 12 typical, legs 18	Kaspar Live #4 [00:06:54]; Dikko [00:13:47]
Eyelid	2 loops per lid crease + ~4 across; 10/10 spans example	Kaspar Live #1 [02:11:22]; Dikko [00:04:16]
Studio mesh	99.9 %+ quads, 0 to 4 valence 6+ (report and justify each), poles ~1.8 % head, ~5 % body, edge CV 0.33 to 0.8	calibration (bx_audit on Blender Studio files)
Auto-remesh hint (not a gate)	edge CV ~0.17 to 0.27 on an unedited uniform remesh, near 0 % closed loops	calibration, sheep
Game	head 2,000 to 3,000 tris, hands ~1,000 both, body 5,000 to 6,000	SpeedChar ep23 [00:09:58], ep24 [00:50:34]
Apply Base levels	2 body, 3 head, hair, clothes	Kaspar Live #5 [02:15:03]

Quality gates

Code (thresholds [added], calibrated on Kaspar's sheep: 0.24 % mean, 47 % inside, coverage p95 0.82 %):

retopo_gate = R.topology_report(retopo)
fit_gate = R.fidelity(retopo, sculpt, levels=2)
gate = {"six_poles_to_justify": retopo_gate["valence_6plus"], "no_rim_poles": retopo_gate["rim_poles"] == 0,
        "quads": retopo_gate["non_quads_visible"] == 0 or retopo_gate["quads_pct"] >= 99.0,
        "subd_fit": fit_gate["mean_pct"] <= 0.3 and 40 <= fit_gate["inside_pct"] <= 60,
        "coverage": fit_gate["cov_p95_pct"] <= 1.0, "loops": R.loop_stats(retopo)}
print(gate)

Edge-length CV is never pass/fail (studio meshes read 0.33 to 0.8); 6-poles are reported and justified, not banned (studio 0 to 4). Also: crease_poles == 0, center_poles == 0 on mirrored faces, R.loops_around(retopo, eye, normal, max_radius) returns at least 3 closed rings per eye and mouth, ring counts equal before every merge, triangles only in hidden or rigid areas, budget met (R.tri_count). Visual (bx_review, same views as the sculpt): cage in wire (rings close, no wandering loops, poles in still areas) and subdivided matcap (lumps, pinching at poles, creases held), then the cage alone: it must read without subdivision (Kaspar Finale [01:58:41]). Score with references/critique.md.

Common mistakes

Mistake	Looks like	Fix
Carving around QuadriFlow's closed eye slits	every eye ring odd (47, 45), no all-quad fix	R.parity_strip between the two odd boundaries BEFORE carving
Counts inherited from a dense cut, or forced far below it	eye rings 46 to 48; or pole density 9.5 %	count=(lo, hi): each reduction unit costs 4 poles; ranges took the sheep from 9.5 to 5.7 %
3D-nearest lookups where surfaces touch	a band or socket cut bites an ear lying on the cheek	geodesic fields, curve_verts, oriented nearest_vertex (built into bands and tubes)
Tight rings for an edge-flow loop	a ridge after subdivision	band spacing = edge length; tight only for creases
Shipping a snapped cage	subdivided surface inside the sculpt (87 %)	R.fit_subdiv
Shrinkwrap offset for volume	cavities worse	fit, offset 0
Raw quadriflow_remesh on a small dense mesh	CANCELED, "needs to be manifold"	R.quadriflow (scale trick, voxel fallback)
QuadriFlow symmetry left at its default True	wrong half on a posed mesh	symmetry=False unless symmetric
Relaxing an auto-remesh	tips and ears shrink	relax hand-built layouts only
Mark Sharp / Seam as reminders	shading breaks (4.1+), unwrap cuts	retopo_crease attribute (P5)
Grid fill on an odd loop	FINISHED, 0 faces	count first; R.grid_fill raises
Connecting parts early	every new loop runs everywhere	islands, count, merge last
Trusting mean error	lumpy fit passes	matcap review, gentle fit
Re-shrinkwrap after Apply Base	volume lost again	never (Kaspar Live #6 [00:20:08])
Game quads left non-planar	lines in the normal map	R.triangulate_twisted, Triangulate on export

Blender 5.2 notes

• Hidden Wire is the Retopology overlay (overlay.show_retopology, offset 0.01 m since 4.5); "Project Individual Elements" is snap_elements_individual = {'FACE_PROJECT'}. GUI only.
• Slide Relax is the Relax Slide brush asset (strokes need a window, bx_gui); headless use R.relax. F2, LoopTools, BSurfaces are extensions; mesh.circularize, space_edge_loops_evenly, flatten are built in; mesh.select_by_pole_count exists (4.4).
• The BCON22 build's Shrinkwrap showed Smooth Factor 0.05 / Repeat 1 (frame 00:22:53); 5.2.1's Shrinkwrap has no such option.
• sharp_edge is always honored since 4.1; Auto Smooth is gone (Smooth by Angle modifier).
• Apply Base = object.multires_base_apply(apply_heuristic=True); QuadriFlow lost its shortcut and has the traps above.

References

• references/procedures.md: tested code, P0 to P14 (inspect, face sets P1b, stack, tubes, geodesic limbs P3b, freeze/relax, annotations, QuadriFlow, map-line loops and rings, sockets, volume, checks, game, expression test, handoff, AI-mesh finishing). Load before writing any code.
• references/expert-notes.md: principles by expert with timestamps, face and body loop maps, clothing, game rules, the sheep experiment in full. Load when planning a face, body, clothing or game mesh.
• references/critique.md: the rubric to judge your own retopo. Load at every review.
• references/sources.md: all 24 source videos with best timestamps.