Skip to content
newblender

12 — Procedural Still Life (Blender 2.93 splash): how an agent could make it
Original Erindale Woodford, 2021, Geometry Nodes procedural art (Blender 2.93 LTS splash) · BlenderNation
Agent-readiness today Today: everything in the piece is a Geometry Nodes (GN) tree plus procedural shaders and Cycles, and the live bridge can build, evaluate, measure and render all of it now. What it cannot supply is the eye: composition, the sunlit mood, which vessel silhouettes are beautiful.
Difficulty 3 (one artist’s piece; a faithful-quality version is about two weeks of agent work plus owner direction)
First slice One GN “vessel generator” (a profile curve revolved, with lip, foot and fluting as inputs) that fills a workbench with 12 seeded, non-intersecting pots, plus GN stems and flowers in two of them, lit by one warm sun through a window and rendered in Cycles.

1. What the original actually is

Section titled “1. What the original actually is”
  • One still image: a sunlit potting-shed workbench with ornate vessels and flowers ready for arranging (index entry 12). It shipped as the Blender 2.93 LTS splash.
  • Fact (BlenderNation): it was made “entirely procedurally using Geometry Nodes” by Erindale Woodford (“nodemancer”), chosen to show off the expanded GN toolset in 2.93.
  • Assets inside the image (interpretation from the picture): bench and wall, a family of turned vessels, pot rims and feet, stems, leaves and petals, small scattered debris, a window light.
  • Follow-up: the “Geometry Nodes Flower Shop” (2022, Blender 3.0) built a whole shop from a single node tree (BlenderNation).
  • The file exists: blender.org’s demo-files page lists “Still Life” by Erindale Woodford under CC-BY, linked to the Blender Cloud gallery. That is a layer-3 reference we can fetch (small file; not downloaded for this note).

2. How the humans made it

Section titled “2. How the humans made it”

Stages: S01 Modeling, S03 Shading, S04 Geometry Nodes, S09 Lighting & rendering.

  • Fact: procedural modelling with GN: instancing, curves, distribution (index entry 12; BlenderNation).
  • Fact (version history): 2.93 GN was the attribute-based system; Fields arrived in 3.0. The 2.93 tree therefore used named attributes passed between nodes, which today’s Fields model replaces with function graphs (dev docs features/nodes/fields.md: a field is an immutable function graph evaluated in a context).
  • Interpretation: each vessel family is a parametric recipe (profile + revolve + displacement); flowers are curves with instanced petals; the bench layout is a distribution with collision avoidance. The art is in tuning those parameters and in the light.

How GN is used in production today (local evidence):

  • S01 census (docs/discovery/s01-modeling/02-census.md): 170 GN node groups and 836 GN modifiers across asset files; 8,423 GN modifiers in shot files. The most-used nodes are assembly utilities (Join Geometry, Realize Instances, Object Info), not “modelling from scratch”.
  • The Blender 4.2 Gold splash ships assets/nodes/procedural_modeling.blend: 16 GN trees, 2,798 faces, a Repeat zone, Sample Curve, Instance on Points and heavy Math/Vector Math use (census JSON). The full splash file carries 1,490 GN modifiers and 77 GN trees (e.g. GN-curve_coral ×68, GN-scatter_strokes ×190).

3. How I would make it: the agent-native plan

Section titled “3. How I would make it: the agent-native plan”
Stage Agent approach Inputs (free sources) Tooling
S04 / S01 Vessels GN group GN-vessel: profile from a Bezier built as data (control points = height/radius table) → Curve to Mesh around a circle (revolve); inputs: height, belly, neck, lip roll, foot, flute count, flute depth, seed Proportions from real pottery reference (owner picks); none downloaded tools/live/bl.py; bpy.data.node_groups.new
S04 Layout GN group GN-bench_scatter: points on the bench top, Poisson-disk via Distribute Points (min distance = max vessel radius), instance seeded vessel variants, Realize None GN Distribute Points on Faces
S04 Flowers GN GN-stem: curve with noise + gravity bend, Curve to Mesh; petals instanced along the stem end by Instance on Points with rotation from Align Rotation to Vector None GN Repeat zone for petal rings
S01 Bench / room Boxes built as data (chair-demo style), Bevel modifier declared None bmesh → Mesh
S03 Shading Glazed ceramic: Principled with coat; colour from a GN-written glaze attribute; terracotta: noise + Voronoi; wood bench from ambientCG/Poly Haven textures Poly Haven, ambientCG (CC0) Shader Attribute node
S09 Lighting One sun through a window cutter, low bounce fill, slight volume for god-rays; camera at 50 mm Optional Poly Haven HDRI (CC0) for window sky Cycles, fixed seed

Build order:

  1. Write checks first (§5).
  2. Bench and back wall as data; camera locked.
  3. GN-vessel with all inputs exposed; render a contact sheet of 16 seeds for the owner.
  4. Owner picks the 3–4 vessel families they like; freeze those seeds as presets.
  5. GN-bench_scatter places 12 vessels; checker confirms no intersections.
  6. GN-stem + petals into two vessels.
  7. Materials; sun and window; test render at 960 px; full render.

4. What works today vs. what needs newblender

Section titled “4. What works today vs. what needs newblender”
  • Works today with the live bridge:
    • All GN nodes listed above exist in 5.2.2 and are creatable from Python; group inputs are settable per modifier.
    • Evaluated geometry is readable for checks (evaluated_get(depsgraph).to_mesh()), exactly as tools/live/checks/chair_checks.py does.
    • Bounding-box and BVH overlap tests (mathutils.bvhtree.BVHTree.overlap) between instances.
    • Cycles render to file and screenshots of the owner’s viewport.
  • Painful today:
    • Authoring a 60-node tree as nodes.new / links.new calls: no text form, no node layout, hard to diff (L-006; nodebpy is the upstream direction).
    • GN gives no result summary: instance counts and bounds must be recomputed by the checker (L-003).
    • Real-world placement wants “this pot sits on the bench” as a declared relation; today it is computed positions (L-009 frames placement as data, which is what we do).
    • Iterating looks pushes no undo from Python (L-008); save a versioned file per accepted look.
  • Not feasible today: nothing mechanical. The infeasible part is knowing which of 1,000 seeds makes the picture sing.

5. Verification plan (RFC 0001 R6, six layers)

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity Every realised vessel mesh is closed and clean 0 non-manifold edges (except declared open rims), 0 degenerate faces, 0 loose verts, normals outward
1 Validity GN trees have no unconnected required inputs and no NodeUndefined 0 (the census found 1,512 undefined nodes in old asset files; ours must have none)
2 Spec Vessel count and variety Exactly 12 vessels; ≥ 4 distinct families; height range 8–40 cm
2 Spec Physical plausibility Every vessel’s lowest point within 1 mm of the bench top; 0 pairwise BVH overlaps between vessels
2 Spec Budget Realised scene < 2 M faces; render at 1920 × 1080 < 10 min on the owner’s GPU
2 Spec Determinism of recipe Same seed → identical vertex hash across two evaluations
3 Reference Compare with Erindale’s CC-BY “Still Life” file (fetch later): vessel count, face counts, light count, camera focal length Report deltas; not a gate on art, a gate on “same kind of scene” (vessel count within ±50%)
4 Downstream Render has no fireflies or black holes < 0.01% pixels > 20× local median; 0 NaN
4 Downstream UVs exist for textured parts 100% of bench/wall meshes with a UV layer
5 Appearance (warning) Vision model compares the render to the 2.93 splash on “sunlit workbench, ornate vessels, flowers” Warn below 7/10
6 Taste (owner) Owner picks vessel families, flower placement, crop Sign-off on the final frame

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • Choosing beautiful silhouettes. A vessel recipe yields thousands of valid pots; maybe 1 in 20 is lovely. That ratio is the human’s.
  • Arrangement. The “ready for arranging” feeling (a fallen stem, one pot slightly askew) is deliberate imperfection; a scatter with collision avoidance looks tidy and lifeless unless directed.
  • Light. The warm window shaft and how much it rakes across glaze is the mood of the piece.
  • When to stop. GN invites endless parameters; the artist knows which ones matter.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): in the owner’s Blender window, a plain bench and back wall appear, then a row of 12 different pots grows on the bench as the GN modifier is added. Changing the seed input reshuffles every pot’s profile and fluting in place; changing flute_count on one preset re-carves it live. Two pots sprout curved stems with ring-instanced petals. A Cycles render with a warm raking sun is saved beside a check report (counts, overlaps, validity) in ~/newblender-data/demos/.
  • Full reproduction: 40–80 agent-hours for the vessel, flower, debris and layout recipes plus checks; < 5 GPU-hours of render; 10–20 human-hours of art direction (seed picking, arrangement, light).

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • Erindale’s actual tree structure is unknown until the CC-BY file is fetched; our recipe may differ in approach even if the picture matches.
  • 2.93-era attribute nodes in that file may load as legacy or undefined nodes in 5.2.2 (compare the census NodeUndefined open question).
  • Vision scoring of “ornate” is weak; expect the owner loop to dominate.

9. Sources

Section titled “9. Sources”