ReSkate Studio compilation

ReSkate Studio’s compile-map consumes map.json and geometry in GLB files, producing Frostbite level content. The measured compiler versions are the older 2026-10-06 build and 20261012; their texture placement differs. Observed

Generated world and residency

Studio automatically streams render cells when a map spans more than 600 m or contains more than three million triangles; --stream and --no-stream override that decision. Its cells are at most 200 m wide, and geometry crossing cell lines is cut. A measured 703-object input becomes 23,419 pieces. Observed

The older compiler places all textures and one depot in the resident root: one measured map has 3,786 root textures and none in its 149 cell bundles. Studio 20261012 writes per-cell depots, with no own root depot when all its materials belong to cells. Observed

Each compile numbers texture names independently and generates new GUIDs. Two independent compiles can therefore reuse names for different partitions or assign different identities to semantically equal data. Observed

The generated world root is 9.6 km. Studio copies a 38-bundle BAM environment closure, including San Van’s sky, sun, fog, exposure, 23 reflection probes and 787 reflection volumes at stock coordinates. It copies BAM agent definitions and places content in referenced LayerData using content bounds rather than directly in quadrant Objects with exact quadrant bounds. Observed

Geometry and input semantics

Studio 20261012 can merge render pieces with identical twelve-number transforms into one prototype named after one input. It emits ObjectReferenceObjectData and a blueprint per placement, rather than StaticModelGroups. Every root-resident mesh needs a root MeshVariationDatabase entry; water section meshes are root-resident. Observed

The GLB importer reads POSITION, NORMAL, TEXCOORD_0/1 and _RESKATE_BLEND, derives tangents from UVs and ignores glTF TANGENT and COLOR_0. Ordinary generated vertices use Float3 position, Half2 UV0 and packed tangent space. Input materials cannot directly select an arbitrary stock material key. Observed

LOD options include --no-lods, --lod-ratios, --lod-distances and --lod-min-triangles. In the measured Numbani map, 299 of 552 MeshSets have three LODs under defaults. There is no corresponding map.json input for supplying a source game’s separate low/medium models as authored LODs. Observed

Blender authoring conventions

Studio also takes maps authored in Blender through its add-on. A third-party editor that writes such files states the conventions below in its exporter’s source; they are its author’s statements, not measurements on the game. Sourced · ReSkate Map Creator

Input	Stated convention	Confidence
Collections	Map, Markers, Grind curves, NPC routes, Lights	Sourced · ReSkate Map Creator
Per-object properties	sk8_collision_mode and sk8_hide_from_pause_map on meshes; curves flagged sk8_grind_curve with a radius and a surface	Sourced · ReSkate Map Creator
Spawn	One empty named spawn with the property xl_marker = "spawn"; the skater faces the empty’s local −Z	Sourced · ReSkate Map Creator
Travel points	Empties whose names start with TravelPoint	Sourced · ReSkate Map Creator
Lights	A seven-bit sk8_light_tod mask: morning 1, noon 2, afternoon 4, evening 8, night 16, weatherday 32, weathernight 64	Sourced · ReSkate Map Creator
UVs	The game reads the mesh’s first UV layer	Sourced · ReSkate Map Creator
Rotation	Rotation belongs on the mesh objects; a rotation on a parent empty is applied about a different axis in the game (yaw becomes pitch or roll)	Sourced · ReSkate Map Creator
Height	The exporter lifts every position by 200 m, stating that the game collides badly with a map at the world origin	Sourced · ReSkate Map Creator

How Studio turns grind curves into collision is under physics and grinds.

Does collision misbehave for a map at the world origin?

The 200 m lift rests only on the editor author’s statement; no measurement of collision near the origin is recorded. Unknown

Material traps

vertex_blend: true on decal materials uses scalar _RESKATE_BLEND values and writes Half4 UV0 with weight in z. Studio patches a decal program to multiply opacity by that weight and appends a private program. The resulting shader-state database has 3,620 records versus stock 3,619. Observed

Studio’s executable describes lifting decal layers by 0.008 * (i+1) metres. Observed A census of compiled map.json levels instead finds decal positions stored verbatim as float32, with no lift, bias or order added; the 8-mm lift appears only in Studio’s Blender importer, for Mix-node decal layers. Observed

Does compile-map lift decal layers for map.json input?

The lift is described from the executable’s strings, while compiled map.json output shows none. Whether the lift applies only to Blender-imported layers, or only with vertex_blend, is not established. Unknown

domain=decal and any alpha=blend material compile to row 1444/3612, and alpha=mask to the cutout row 825/1891/3096; Studio’s opaque row is (682, 819, 1474, 2756), whose depth and shadow copy section carries the prepass 3560 and shadow program 1262. The stock properties of these programs are listed under decals. Observed

ReSkate’s stock-program-only merge can remove that private program. Measured outcomes include boot faults, invisible overlays on builds that boot, and a separate unresolved preparation fault. See the crash catalogue.

In SkateBIRD level12, alpha-blend materials compile to rows 1444/3612 and disappear as world geometry, while alpha-tested materials using 825/1891/3096 render. This observation does not establish that every stock blend program is unusable. Observed

Default flat-normal and mask textures repeated across streamed cells cause a session-persistent glow failure. Their working residency is one root copy. Splitting transparent layers at cell boundaries also changes their bounds, which can change overlay draw order. Observed

Other compiler boundaries

Feature	Measured behaviour	Confidence
Travel points	At most 25 per compile	Observed
Pause map	--pause-map 2d uses ui/map_image.png, bounds in map_image.json, and map_clouds.png	Observed
initfs_win32	Output copy equals stock Data/initfs_Win32; inspected ReSkate versions do not read the mod copy	Observed
GI-grid bake	Hard-disabled in 20261012, independently of --no-gi	Observed
Enlighten bake	--enlighten can emit a real database	Observed
Custom colour shaders	Cosmetic-only function of UV, base colour, clock and four channels; no world-position, normal or vertex-offset input	Observed

The emission path and GI limits are detailed under lighting; collision grouping has separate thresholds.

Measured compiler memory grows at about 0.263 GB per million compiled triangles: 54.1 million uses 14.1 GB, and 73.0 million uses 19.2 GB. These are observations of that input family, not a hard triangle-format limit. One measured tile spends about half its 142-second compile in packaging. Observed

Two compiles of one export differ in 201,352 of 506,104 manifest rows because identities are regenerated. Raw binary differences between such compiles therefore do not independently demonstrate changed source content. Observed Compiling one map twice from a byte-identical input folder with 20261012 gives different bytes in the default install package’s cas_01.cas (33,973,484 and 33,973,268 bytes) and cas_02.cas, the initial install package’s cas_01.cas, globals.toc and the level’s .toc; dingolevel_root.toc, initfs_win32 and layout.toc are identical, and reskate-build.json differs only in its time stamp. Observed · two compiles of one map Whether a change of input changed a level is shown by the input Studio reads, not by its output. Inferred