Scenario Blender Retopology
ASecurityUse when turning a dense sculpt, scan or AI-generated mesh into clean topology in Blender, for animation, subdivision or a game low poly with a triangle budget. Also when choosing between QuadriFlow or voxel remesh and manual retopo, placing loops around eyes, mouth and joints, deciding pole placement, setting up a Shrinkwrap cage, fixing volume loss after subdivision, preparing a low poly for baking, or auditing topology for spirals, 6-poles and triangles. Keywords: retopo, retopology, quad ...
- 871 stars
- 0 votes
- 1 copy
- 3 views
- Added September 26, 2026
Works with
Security analysis
100/100Pro scans all 5 files and shows the line behind each finding
npx -y skills add scenario-labs/skills --skill scenario-blender-retopology --agent claude-codeAre you the author of Scenario Blender Retopology?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/scenario-labs-scenario-blender-retopology)---
name: scenario-blender-retopology
description: "Use when turning a dense sculpt, scan or AI-generated mesh into clean topology in Blender, for animation, subdivision or a game low poly with a triangle budget. Also when choosing between QuadriFlow or voxel remesh and manual retopo, placing loops around eyes, mouth and joints, deciding pole placement, setting up a Shrinkwrap cage, fixing volume loss after subdivision, preparing a low poly for baking, or auditing topology for spirals, 6-poles and triangles. Keywords: retopo, retopology, quad remesh, clean up an AI mesh, edge flow, low poly, bake prep."
license: MIT
---
# 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`](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 %):
```python
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`](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`](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`](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`](references/sources.md): all 24 source videos with best timestamps.
Files in this skill
- SKILL.md
- references/critique.md
- references/expert-notes.md
- references/procedures.md
- references/sources.md
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.