Skip to content
newblender

15 — Park Lipence: how an agent could make it
Original Polygoniq, 2021, architectural visualisation / photoreal landscape (Cycles) · Blender 3D Architect
Agent-readiness today Partial: terrain, paths and park architecture as data, Geometry Nodes (GN) scattering with masks, physical sky and Cycles all work through the live bridge today. The bottleneck is photoreal vegetation: the original used Polygoniq’s commercial botaniq library, and free CC0 plant libraries are thin, so the hero trees either need a generator (Sapling, GN tree demo) or a paid library. Photoreal “believability” is also owner taste.
Difficulty 4 (a small studio’s showcase set; a convincing park section is about two weeks of agent work, mostly vegetation quality and render tuning)
First slice A 40 × 40 m park patch: gently rolling terrain built as data, a curving gravel path that masks the scatter, GN-scattered grass clumps, shrubs and 8 generated trees, a simple timber pavilion built with declared modifiers, Blender’s physical sky at late afternoon, and an eye-level Cycles render.

1. What the original actually is

Section titled “1. What the original actually is”
  • A set of photoreal stills of a park near Prague (Lipence is on the city’s south-western edge), with “open spaces, vast vegetation, and unique architectural designs” (Blender 3D Architect, fetched 2026-10-07). The number of images and whether it was commissioned or self-initiated are not stated.
  • A vendor showcase: Polygoniq makes the add-ons used: botaniq (vegetation), materialiq (materials), traffiq (vehicles). The article appeared with a Blender Market discount in August 2021.
  • Deliverables (interpretation): terrain and lawns, a path network, park buildings/pavilions and furniture, thousands of plant instances at several scales (grass, shrubs, trees), parked or passing vehicles at the edges, a sun-and-sky lighting setup, and colour-graded stills.

2. How the humans made it

Section titled “2. How the humans made it”

Stages: S01 Modeling, S03 Shading, S04 Geometry Nodes, S06 Layout, S09 Lighting & rendering.

  • Fact: Cycles rendering; botaniq, materialiq and traffiq libraries (article).
  • Fact (index): architectural modelling, vegetation scattering, library materials, photoreal Cycles lighting.
  • Interpretation: the architecture is modelled from plans with standard archviz modelling (extrude, bevel, array); vegetation is scattered by the library’s scatter tools (botaniq is built on particle systems / GN scatter with LODs) and art-directed by hand painting density masks; hero trees are placed one by one for composition; lighting is a physical sky or HDRI; post grading in the compositor or an external tool.
  • Local evidence gap: the S01 census has no archviz corpus yet (brief §8: “add a hard-surface / archviz set”; census next steps name Barcelona Pavilion and Classroom). This file’s plan is therefore not grounded in a measured archviz census.

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
Site data Park outline, paths and building footprints from map data; terrain from an elevation model or a declared noise if none OpenStreetMap (ODbL, not CC0: attribution required); Prague open geodata portals for elevation (unverified availability) Python parse of an OSM extract via tools/live/bl.py
S01 Terrain Grid mesh built as data, heights from the DEM or noise; path_mask attribute written by distance to path curves Same bmesh / GN Set Position + Geometry Proximity
S01 Architecture Pavilion, benches, lamps as data with real dimensions; Bevel/Array/Solidify declared (chair-demo method) Owner’s reference photo or plan bmesh + declared modifiers
S04 Scatter GN GN-park_scatter: Distribute Points on Faces with density = (1 − path_mask) × slope falloff × noise; three layers (grass clumps, shrubs, trees) with min-distance per layer; collection instances; camera-distance culling and LOD switch Poly Haven models (CC0 plants, rocks, a few shrubs); Blender GN demo “Tree, leaves and grass” by Simon Thommes (CC0); Sapling Tree Gen extension (free) for tree variety GN
S03 Materials Lawn, gravel, timber, concrete from CC0 PBR sets; leaf translucency in Principled; per-instance hue jitter from GN Random Value ambientCG, Poly Haven textures (CC0) Shader Attribute (instancer)
S06 Layout Cameras at 1.6 m eye height along the path; hero trees placed as data at owner-chosen positions None Python
S09 Lighting Sky Texture (physical sun/sky) at a declared date/time/latitude (Prague ≈ 50.0° N), or a CC0 HDRI Poly Haven HDRIs (CC0) Cycles; adaptive sampling; denoise
S10 Grade Compositor: mild filmic/AgX contrast, slight glare None Compositor via Python

Build order:

  1. Write checks (§5); decide the site (real OSM extract or a declared synthetic plan).
  2. Terrain and path curves; path_mask attribute; check path width and slope.
  3. Pavilion and furniture as data with declared modifiers.
  4. Scatter layers one at a time, viewport in solid mode with instance display as bounds to stay interactive.
  5. Trees: generate 6–10 variants (Sapling or the GN tree demo), place 8 hero trees as data, scatter background trees.
  6. Materials, sky, camera; 25% test render; owner review; final 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:
    • Terrain, paths and buildings as data with declared modifiers and spec checks, as in demo 001 (12/12 checks, ~/newblender-data/demos/001-chair/report.md).
    • Building and parameterising GN scatter trees from Python; evaluating instance counts and positions for checks.
    • Sky Texture, Cycles settings, compositor nodes and render-to-file from Python; screenshots of the owner’s window (tools/live/snap.py).
  • Painful today:
    • Scatter art direction is normally done by painting density weights; that is a brush stroke stream (L-009). We replace it with declared masks (distance to path, slope, noise, owner-placed exclusion curves).
    • Hundreds of thousands of instances make the live viewport slow; the right home for final evaluation and render is a headless profile (L-011).
    • GN returns no summary of what it scattered; counts, densities and minimum spacing are recomputed by the checker (L-003).
    • Free photoreal trees: Poly Haven’s CC0 plant set is small, Sapling trees read as “generated” up close; this is an asset-supply problem, not an engine one.
  • Not feasible today:
    • Matching botaniq’s hero-tree realism with only CC0 assets. A fair reproduction either buys the library (owner decision) or accepts a lower bar for close-up trees.
    • Traffiq-quality vehicles from CC0 sources at hero distance.

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

Section titled “5. Verification plan (RFC 0001 R6, six layers)”
Layer Check Threshold
1 Validity Terrain and architecture meshes clean 0 non-manifold edges on closed parts; 0 degenerate faces; normals up on terrain (100% faces with normal·Z > 0)
1 Validity Every instanced collection and texture resolves 0 missing objects or images
2 Spec Patch dimensions and terrain 40 × 40 m ± 0.1 m; max slope < 15° on lawn areas
2 Spec Paths stay clear 0 scattered instances with origin inside a path (path_mask > 0.5); path width 2.5–3.5 m
2 Spec Scatter density per layer Grass ≥ 20 clumps/m² on lawn; shrubs 0.2–0.5 /m² away from paths; exactly 8 hero trees, ≥ 4 m apart
2 Spec Grounding Every instance base within 3 cm of terrain surface (raycast)
2 Spec Architecture Pavilion footprint and height match the declared plan within 1 cm; scale applied; Studio naming GEO-park_*
2 Spec Budget 1920 × 1080 Cycles render < 30 min on the owner’s GPU; peak memory within GPU VRAM
3 Reference Sun direction matches the declared date/time/latitude Sun azimuth and elevation within 1° of an ephemeris calculation
3 Reference (Later) archviz counter-corpus: compare instance densities and light setup with the CC0 Classroom and CC-BY Barcelona Pavilion demo files Report deltas; not a gate
4 Downstream Render health 0 NaN; fireflies < 0.01% pixels > 20× local median; sky not clipped (< 0.5% pixels > 1.0 after view transform)
4 Downstream Determinism Same scene, seed, frame rendered twice: PSNR ≥ 45 dB
5 Appearance (warning) Vision model compares to the Park Lipence images on “photoreal park, believable vegetation, no CG tells” and lists tells (tiling, repeated trees, floating plants) Warn below 7/10 or any listed tell
6 Taste (owner) Owner judges realism, composition and season/mood Sign-off; notes become mask, density and light parameters

6. Where the human is still needed

Section titled “6. Where the human is still needed”
  • “Does it look real?” Photoreal is judged by the eye against lived experience: grass colour variation, how leaves catch the light, the absence of repetition. Machines can flag tells, not sign off realism.
  • Composition of nature. Where trees frame the building, where an open lawn breathes: archviz is selling a place, and that is a design decision.
  • Asset budget. Whether to buy a vegetation library (botaniq or similar) is an owner decision with money attached.
  • The architecture itself. If it represents a real design, the architect’s drawings are the source; an agent should not invent a real building.

7. Effort

Section titled “7. Effort”
  • First slice (live demo, ~1 day): in the owner’s Blender window, a flat plane rises into gently rolling ground; a gravel path curves through it. Adding the scatter modifier, grass fills the lawn but stops cleanly at the path edges; shrubs and eight trees appear; changing grass_density or moving a path control point re-scatters live. A timber pavilion with bevelled beams sits at the path bend. The viewport switches to rendered mode under a late-afternoon physical sky with long shadows. A Cycles still and a check report (density per layer, path clearance, grounding, sun angle) land in ~/newblender-data/demos/.
  • Full reproduction: 60–120 agent-hours for site, architecture, scatter system, cameras and checks; 5–20 GPU-hours of final stills; 20–50 human-hours of direction and realism review; plus the cost of a vegetation library if the owner chooses one.

8. Risks and unknowns

Section titled “8. Risks and unknowns”
  • The original’s exact site, image count and commission status are unknown; the article is short and promotional.
  • OSM data is ODbL (attribution and share-alike on the database), not CC0; check before publishing derived renders with the data.
  • Free trees may cap realism; this could make the honest result “good archviz, not Polygoniq-level”.
  • GPU memory with many unique tree meshes; instancing must stay instanced until render (do not Realize Instances on the scatter).
  • No archviz census exists yet, so modelling norms for this genre are assumed, not measured.

9. Sources

Section titled “9. Sources”