Scenario Unreal Materials

Unreal materials (master materials, Substrate, textures, shader cost)

Expert level means a few master materials whose instances cover the kit, a known and budgeted permutation count, cost proven in milliseconds on a fixed camera, and physically plausible values before any artistic push. The graph is authored once (scripted build or template); everything else is instances, Custom Primitive Data and textures that follow a role policy. Never trust the instruction count or a pretty preview. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-unreal-expert (channels, review loop, 5.8 traps).

Status (2026-09-24): Unreal is not installed; offline tests pass (tests/code/unreal-materials/run_all.sh, editor layer on a fake unreal). Every snippet is not yet run in Unreal: run P0 first.

Stance (the expert delta)

• Permutations = usage flags x unique static parameter sets across instances (x quality levels with a Quality Switch), never a node count. A parent flag adds a shader row to every unique child (Lauf and Nadro, Fortnite [wobQ8ZKQpbc 00:24:25, 00:05:32]); parents carry the fewest flags, instances opt in (5.8 Usage Flag Overrides [verify]). Before adding a switch, count the instances that would set it (switch_ranking; Fortnite kept its top 10 [00:29:31]).
• A static switch must gate texture samples, or heavy math most instances skip. Anything else is an If (parameter into A, 0.5 into B, A > B pin, never A == B), an Enum-bound scalar or a Channel Mask Parameter; the Static Component Mask Parameter is obsolete (Lauf [00:40:24, 00:40:56, 00:42:35]). An editable switch at default still sets the HSPR tag (5.7).
• The instruction count is not cost. Sine plus cosine read +2 but cost about 40 GCN cycles; a fetch waits about 200 (Swierad [y0QASid1v8w 00:20:57, 00:35:57]). Math independent of the fetch hides inside that wait (a screen's scanlines are nearly free); math on the sampled value is paid in full [00:39:34, 00:46:03]. Budget samples and transcendentals (cost_census), open every material function ("cheap" in a name is not a measurement [00:22:31]), decide on ms at a fixed camera.
• Substrate starts from one slab (Morgan, Epic [SqPaL8HS_Lw 00:25:45]): a Slab or the Metalness helper, never a hand-added Substrate Shading Model node (the conversion target, Substrate overview). Blendable is a ceiling for every platform, SM6 included (5.7 article); Adaptive falls back per platform by itself and costs 20 to 80 B/px (663 MB at 4K), about +15% cook and forced DBuffer decals (Ben Cloward [P5I38f2O6W8 00:05:46, 00:06:54]). Converting a material is one-way.
• Metal next to paint is two slabs, not gray metallic (gray reads as plaster, Ben [VrY_SSvWdQ4 00:05:39]). With Use Parameter Blending a Horizontal Blend is near-lossless and one slab everywhere (Morgan [00:37:35]): mask sharpness is then artistic, not cost. "Mix exactly 0 or 1 evaluates one slab" holds only without parameter blending, on Adaptive [00:29:38]. Parameter blending on a Vertical Coat is lossy and propagates beneath: plan it early [00:38:08, 00:38:43].
• Detail comes from packing and tiling, not resolution. Two samples per terrain layer (CR + NOH), a 2K game ceiling, compression by role, sRGB only on color, packed normals never as Normalmap, BC7 for unrelated channels (Ben [-UZlUUQSGgQ 00:12:36], [gjOO5g4cgng 00:11:02]; Sumo [SAr7oPKsgLE 00:29:20]; textures doc). Memory truth is the Platforms panel and listtextures in a cooked build; RDO shrinks disk, not memory.
• Glass is matter, not opacity: Colored Transmittance only when tinted (Gray is cheaper), Surface Translucency Volume, Pixel Normal Offset for flat panes, slab Simple Volume with Transmittance-To-MFP (Ben [sf-K257zWh8 00:01:39, 00:09:17]).
• An RVT is a cache captured orthographically at time 0: no Time, Panner, camera distance, parallax or MPC in the write pass; a page change redraws all its writers; each distinct RVT sampled has a large fixed cost (PrismaticaDev [RLEPA16QDRw 00:20:14, 00:11:59, 01:04:29]).

Establish first

Workflow

0. Probe and decide. P0 um.probe() lists what 5.8 really exposes; P1 reads the ini and runs um.choose_gbuffer_format(targets, needs), whose reasons state the ceiling and the fallback. GATE: CANDIDATES corrected; the format decision written.
1. Plan the family. One master per domain and blend mode (surface, glass, screen, decals, UI Material, landscape), each with its switches and what they gate, approved presets, usage flags per mesh type, sample and closure budgets, MPCs. GATE: every switch gates a texture and has a switch_ranking count; channel picks are Channel Mask Parameters; architecture written.
2. Textures (P2): pack_channels, import_textures, audit_textures(apply=True), albedo_stats. GATE: audit passes; resident size at most 2048; lint_graph(g, textures=...) finds no sampler mismatch; normals keep Normalize after making Mips; albedo sRGB 20 (50 rough) to 240, metals at least 180.
3. Build masters (P3, P4): um.build_kit_surface or a template: texture-gating switches, an If for the paint-mask source, Custom Primitive Data for per-actor wear and dirt, Specular = 0.5 x (1 - cavity) as specular occlusion. Substrate: the Metalness helper feeds a base slab, a bare-metal slab takes albedo 0 and a measured F0, a parameter-blended Horizontal Blend takes the wear mask. Options: value-clamped ID masks, 4 bands per channel (Sumo [00:14:54]); world-aligned texture and normal across modular seams, 3 samples per texture (Ben [pXOknekvmwE 00:12:57]); world-space directional dust (Sumo [00:20:11]). GATE: no failed links; lint_graph free of fail and warn; one closure; clean compile log.
4. Instances (P5, P6, P16): only make_presets sets static switches; children change values; Custom Primitive Data varies placed actors. GATE: unique sets at most the preset count, no override at default, HSPR only on presets; audit_instance_values_project clean (dielectric F0 at most 0.08, metal albedo 0, Metallic 0 or 1).
5. Usage flags and shaders (P7): pin needs on instances, then clear the parent; auto usage off; Used with Static Mesh off on decal, UI and Niagara-only materials (Nadro [00:18:38]); shader_count per root (an upper bound). 5.8 compiles only connected, non-zero outputs (r.Material.UseShaderCompilationParameters): deleting nodes to strip shaders is folklore. Skin-cache and Slate levers (P7) only after measuring. GATE: shaders within budget; the fallback board shows no gray default material.
6. Special families. Screens: emissive calibrated with the lighter; TSR smearing on scrolling text gets the experimental Temporal Responsiveness node, masked, with r.Velocity.TemporalResponsiveness.Supported=1 and a small WPO on Nanite meshes (5.7 article). Decals (P14) end in Convert To Decal; overlapping DBuffer-expression decals do not blend; each receiver's Decal Response covers only the channels its decals write (decal_response_for). Landscape P11, stylized P12 and P13, switch triage P8. GATE: references/critique.md for the family.
7. Cost A/B (P9, live editor): fixed camera, r.ScreenPercentage 100, warm-up, CSV profile, medians of GPU and pass time, cost_census beside them. GATE: deltas within budget; instruction counts recorded as context only.
8. Look review (P10, P15): preset board under the level lighting, buffer visualizations, Substrate views. For a Substrate surprise, open Window > Substrate first: features in downgrade order, parameter-blending preview (Morgan [00:12:51, 00:38:43]). ue_review.review_images runs before anyone looks. GATE: critique rubric passes.
9. Cooked check and handoff (P17): listtextures, stat streaming and r.Streaming.PoolSize through texture_memory_findings; hand over reports, captures, open [verify] items.