env-lab — finding a way to build beautiful game environments with an LLM in the loop

Thesis

Beauty in an environment is roughly asset quality × density × lighting. An LLM is bad at the first when asked to model, weak at the second when asked to place instances one at a time, and good at the third. So the tooling never asks the model to make art. It directs three factories and judges the result through a reference-anchored critic:

Factory	What the model authors	What produces the pixels
Asset	prompts, picks, dimensions, families	image-gen → image-to-3D → Blender cleanup; kitbash packs; proper procedural generators (displaced icospheres, not tapered cubes)
Density	terrain design params, biome/scatter rules	heightmap + erosion (tools/terrain), rules-based scatter into instanced meshes (tools/ue/ue_e1_scatter.py)
Look	lighting, fog, grade, post stack, camera	Unreal (Lumen, PPV) — the look-round method from AttunementDemo, unchanged

The critic scores stills against a reference board you pick (references/), on a fixed checklist (critic/RUBRIC.md). Your ratings accumulate in critic/ratings.csv so the critic can be checked against you. The model never canonizes a look. Only you do.

Why this exists

Documents/AttunementDemo (UE 5.8) proved the pipeline — remote Python into the editor, headless Blender, screenshot look-rounds — and produced stills that were still empty planes with floating slabs. See the diagnosis in experiments/01-barrow-density/EXPERIMENT.md.

Layout

experiments/<nn>-<name>/   one hypothesis each: EXPERIMENT.md, params, shot list, results
tools/terrain/             heightmap.py — macro shape + fBm + droplet erosion → npy/png/meta
tools/blender/             headless generators: gen_terrain.py (heightmap→FBX), gen_rocks.py (scatter kit)
tools/ue/                  scripts that run *inside* Unreal via remote.py: setup, scatter, snap
critic/                    RUBRIC.md, ratings.csv (your judgments), scoring notes
references/                your reference board (frames you love) + BOARD.md
build/                     generated: heightmaps, FBX, snaps (not committed)
assets/                    downloaded kits + SOURCES.md with licenses (not committed)

Running experiment 01

Prereqs: UE 5.8 editor open on AttunementDemo.uproject (remote execution is on), Blender 5.2, py -3 with numpy + Pillow.

py -3 tools/run_e1.py --all          # terrain → kit → UE setup → scatter → snaps
py -3 tools/run_e1.py --snap         # just re-capture the fixed shot list

Stills land in build/snaps/e1/. Compare against build/snaps/baseline/ (the LR10 set).