Framework vs. cold prompts

At a glance

The scores come from one kit: ten automatic checks against level.json, plus one judged by eye. The walker is a person-sized CharacterController (radius 0.35 m, height 1.8 m, step 0.3 m). Costs are at API list price; both cold runs actually ran on a subscription.

Check by check

The same kit and the same level.json scored all three grid builds. Check 10 is a walk, not a line from the brief.

What the kit saw

Each map is the walkable NavMesh at the walker's size, seen from above, with the walked routes drawn on it. The grid is the brief's 4 × 4 layout of 45 m cells.

E entry hall · G guard room · P prison · V vault · K kitchen · D great hall · A armory · T throne room · L library · S study · F forge

Built from what

The cold runs placed pieces one at a time, straight into the scene. Ours places whole rooms, which nest set pieces, which nest the Synty pieces. So ours puts down far fewer objects but draws on a wider range of pieces.

Side by side

Every row is measured with the same probe, run in batch mode on each saved scene; no scene was saved afterwards. The patterns level covers 300 × 400 m, so its totals are not like-for-like with the three 45 m-grid builds.

Which packs

Where each build's pieces come from. Fable leaned on the base Dungeon pack. Opus and our rooms drew about two-fifths of their variety from Dungeon Realms. Our pattern tiles use the base pack, plus a few Dungeon Map pieces that their new set pieces bring. "Generic" is Synty's PolygonGeneric, which the cold-run template also shipped.

Same pieces, different hands

How many Synty pieces each pair of builds has in common. The diagonal is each build's own count. Shading is the overlap as a share of the smaller build in the pair. Two-thirds of Fable's pieces appear in our rooms level too; the overlap with Opus is about the same.

The most-placed pieces

Each build's six most-copied assets. Every build is mostly walls, floors and ceilings. Fable's number one is the Unity cube it used for solid rock; our rooms' is the flame effect on their torches, which draws particles, not triangles.

Names without a tag are from PolygonDungeon, with the SM_Env_ or SM_Prop_ prefix dropped.

Inside the rooms level

The eleven room prefabs, largest first. Each was chosen from the pre-built themed-room library, not made for this level; the reason for each pick is in level.picks.json.

The patterns level nests six d-20 pattern tiles, each used once: flanking route (×2), gallery, hidden cache, sniper location and stronghold. Since 13 Oct the tiles carry 123 different set pieces (185 placed), so its renderers sit up to five prefabs deep.

Read with care

• Flat and nested builds count prefabs differently. A cold run's outermost-prefab count is its piece count; ours is not. To compare variety, use the different-pieces, leaf-asset or mesh rows.
• The patterns level is a different size (300 × 400 m against eleven 45 m cells), so its renderer, triangle and collider totals are not like-for-like.
• Scene file size understates ours. Our content lives in prefab files; the "level data on disk" row adds them back. Those room and set-piece prefabs come from pre-built libraries and were not made for this level.
• Synty textures are atlases, so every build has few materials and textures. Those rows show which packs were used more than how varied the look is.
• Nothing is baked. No build has lightmaps or probes. The patterns level has 96 point lights for its 318 torch meshes, so most of its torches cast no light. The cold runs each add one post-processing volume; ours add one global grade and a colour-mood volume per room (11 in the rooms level, 71 in the patterns level).
• Not measured: draw calls and frame time.
• The framework is not counted as code for ours. It is MacroTileBuilder, PrefabSynthesis and the room and set-piece libraries, all of which existed before this brief.
• The brief favours the framework. It was written to what the framework already has, so treat this as a showcase brief, not a neutral benchmark.