S01 Pass 2 — Dissect: what real production files actually contain

Note

Generated from docs/discovery/s01-modeling/02-census.md.

• Date: 2026-10-07 · Blender: 5.2.2 LTS, headless, --factory-startup --disable-autoexec (embedded scripts and Python drivers never ran; the logs show blocked PyDrivers, as intended)
• Tools: tools/census/census.py (per file) → tools/census/aggregate.py
• Raw data: ~/newblender-data/census/*.json, ~/newblender-data/census_summary.json

Corpus (78 files, all free downloads)

Group	Files	What
Assets (authored data)	51	Sprite Fright shot library (5 chars, 9 env sets, 13 props, node libs, camera rig, light templates), Gold splash asset pipeline (published chars, corals, kelp, FX, node toolkits incl. procedural_modeling.blend), Human Base Meshes v1.4.1, Cube Diorama, VDM brush demo, 3 sculpt demos, cloth-brush demo
Shots (assembly)	27	Sprite Fright 030_0020_A anim/A/B/C/lighting files, splash scenes 4.2 / 4.3 / 4.5 / 5.1 / 5.2

Counting rules: topology counts only meshes local to their file, so linked data is never counted twice. Modifier counts are per file and include linked objects.

Bias warning:

• The sample is dominated by Blender Studio stylised animated films: organic forms, subdivision workflows.
• Hard-surface, archviz, game and CAD-like modelling are under-represented.
• Treat the numbers as “how this studio models”, not “how everyone models”.

Findings (asset files; facts)

Measure	Value	Plain meaning
Local meshes	1,403 meshes, 1.19M faces
Face mix	94.5% quads, 5.4% tris, 0.1% ngons	Production topology is almost all quads. Only 13.5% of meshes contain any ngon at all.
UVs	92.2% of meshes have UV layers	UV unwrapping is part of nearly every asset (S02 matters to modelling)
Shape keys	46 meshes	Mostly character faces/correctives
Median mesh size	111 vertices	Most meshes are small parts; detail comes from modifiers
Modifiers	9,317 on asset objects	Non-destructive stacks are the norm, not the exception
Top modifiers	SUBSURF 5,003 (54%) · DISPLACE 920 · NODES (Geometry Nodes) 836 · SIMPLE_DEFORM 741 · SMOOTH 416 · SOLIDIFY 399 · VERTEX_WEIGHT_MIX 191 · ARMATURE 137 · MULTIRES 91 · LATTICE 90 · MIRROR 80	Low-poly cage + live subdivision is the dominant modelling pattern
Rare modifiers	BEVEL 59 · DECIMATE 14 · ARRAY 9 · BOOLEAN 6 · WELD 1	Hard-surface modifiers are rare in this corpus (see bias)
Stack depth	1 modifier: 3,584 objects · 2: 1,265 · 3: 552 · 4+: 315 (max 15)	Mostly shallow stacks, with a long tail
Node groups	170 Geometry Nodes · 650 shader · 3 compositor	GN is already a modelling tool in production
GN nodes used most	Join Geometry, Realize Instances, Object Info, Set Shade Smooth, Collection Info, Boolean Math, Edge Angle	Assembly and shading-prep utilities more than “procedural modelling from scratch”
Unresolved nodes	NodeUndefined 1,512 in assets, 1,094 in shots	Node types this 5.2.2 build doesn’t recognise (add-on or version drift). Open question.
Linking	162 library references; 5,996 linked objects; 127 overrides	Assets are themselves assembled from other assets

Findings (shot files; facts)

Measure	Value	Plain meaning
Overrides	9,266 library overrides	Shot assembly is override-heavy: the redesign target from 03c S06
Linked objects	16,946	Shots are mostly links, not authored geometry
Modifiers	35,275: SUBSURF 9,248 · NODES 8,423 · LATTICE 3,722 · ARMATURE 3,553 · HOOK 2,078 · CORRECTIVE_SMOOTH 1,350	Deformation and GN evaluation dominate at shot time
GN trees	651; most common nodes: Switch, Capture Attribute, Compare, Field on Domain, Position, Set Position, Random Value, Named Attribute	Field-based attribute logic is the production idiom
Deep stacks	123 objects with 14 modifiers, 124 with 15	Character rigs with long deform stacks

What this means for the workflow script (Pass 1) — interpretation

1. Modelling output is a live recipe, not a final mesh [DF3]. 91% of mesh objects in asset files carry modifiers. The authored artefact is a small cage plus a stack of declared operations. That is already close to an agent-native “declare the outcome” form. The destructive edit-mode steps happen mostly on the cage.
2. Topology rules are strict and checkable [DF4]. A 94.5% quad / 0.1% ngon outcome means the human “eyeball the wireframe” check can become a measured gate.
3. Geometry Nodes is already production modelling infrastructure in both assets and shots. An agent-native modelling layer should treat GN trees as first-class output, not an afterthought.
4. Hard-surface ops (bevel, boolean, array) are rare here. That reflects the studio’s style, not a general truth. A hard-surface corpus is needed before any verdict on those capabilities.

Answers to Pass 1’s open questions

Pass 1 question	Answer
Modifiers live vs applied?	Live: 91% of mesh objects carry modifiers in asset files
Ngon/tri ratio	0.1% ngons, 5.4% tris
Mirror usage	80 Mirror modifiers (assets), 309 (shots)
Shape keys	46 meshes (assets), 224 (shots)
Multires kept?	91 Multires modifiers (assets)
Leftover booleans	6 Boolean modifiers (assets)
Applied scale, duplicate vertices, flipped normals, naming compliance, UDIM count, pivot Empties	Not yet measured. Needs census v2.

Gaps and next steps

• Explain NodeUndefined: add-on node types, removed nodes, or version drift in 4.2–4.5 files opened in 5.2.2.
• Add a hard-surface / archviz corpus: Cycles demos (Classroom, Barcelona Pavilion), benchmark scenes, the GN demo set.
• Per-object census of edit-mode history is impossible. .blend files store results, not the steps that made them. The human steps come only from Pass 1 docs and (optionally) live GUI sessions.