Measure what one frame of a subject commits the renderer to.
Everything here is exact or an explicit upper bound, and the difference is
stated per metric rather than left to a reader:
Counts of nodes, lights, slots, chunks and materials are exact;
Texture allocations use the conservative union of sampler, environment
and shadow consumers. Uncertified identities/profiles or unsafe counts
remain null rather than an invented exact native allocation population;
Triangles, vertices and draw calls for instanced sets are UPPER BOUNDS, taken
over the most expensive level of detail and the most expensive prototype a
slot could select, because level-of-detail selection and frustum culling
are camera facts and a budget must hold for every camera;
Texture bytes are conservative uncompressed-storage estimates from decoded
facts and required target profiles. Unresolved dimensions, representation
or residency facts remain null rather than invented storage.
The upper-bound direction is the only safe one. A budget checked against an
average would pass a production that stutters whenever the camera turns
toward the crowd, which is exactly the frame anyone would have wanted the
budget to catch.
Simulated drawables — cloth panels, planting clusters and water surfaces —
are measured beside the staged ones. They are held by no scene node, so
leaving them out was never "not yet supported": it was a triangle count for a
room the curtain, the fern bed and the pond are missing from, checked against
a budget and reported as cleared. Their cost comes from the domain record
alone and no solve has to run, which is what lets a production be refused
before the first step is integrated.
Owners are the semantic ids of the mask, so a cost in the report and a colour
in the mask name the same thing. Shared resources that draw no pixels of
their own carry a material:, texture: or light: identity instead.
A row that is one further whole-frame pass over geometry the other rows
already counted, such as a shadow view's depth pass or the outline guide pass,
is marked pass rather than own. The pass belongs in the conservative
one-frame peak and its own identity is worth keeping, but its cost is the sum
of every drawable that preceded it rather than the owner's own complexity, so
the report keeps its attribution without ranking it as an editable owner.
Measure what one frame of a subject commits the renderer to.
Everything here is exact or an explicit upper bound, and the difference is stated per metric rather than left to a reader:
The upper-bound direction is the only safe one. A budget checked against an average would pass a production that stutters whenever the camera turns toward the crowd, which is exactly the frame anyone would have wanted the budget to catch.
Simulated drawables — cloth panels, planting clusters and water surfaces — are measured beside the staged ones. They are held by no scene node, so leaving them out was never "not yet supported": it was a triangle count for a room the curtain, the fern bed and the pond are missing from, checked against a budget and reported as cleared. Their cost comes from the domain record alone and no solve has to run, which is what lets a production be refused before the first step is integrated.
Owners are the semantic ids of the mask, so a cost in the report and a colour in the mask name the same thing. Shared resources that draw no pixels of their own carry a
material:,texture:orlight:identity instead.A row that is one further whole-frame pass over geometry the other rows already counted, such as a shadow view's depth pass or the outline guide pass, is marked
passrather thanown. The pass belongs in the conservative one-frame peak and its own identity is worth keeping, but its cost is the sum of every drawable that preceded it rather than the owner's own complexity, so the report keeps its attribution without ranking it as an editable owner.