Rendering and lighting (lighter / graphics TA)

Expert lighting in Unity is a measured loop: fix exposure, light to anchors, bake, grade, judging every stage from the same camera bookmarks with numbers (gray cards, clipping, bake memory, native passes) before taste. This skill drives URP 17.3 on Unity 6.3 from code; everything here ran in Unity 6000.3.21f1 on an Apple M5 Max (Metal) on 2026-09-24. 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-unity-expert (channels, ut_run, AgentCapture, AgentProfile, ut_review, 6.3 traps).

Stance (the expert delta)

1. Exposure first, lights second (Pierre Yves Donzallaz, yqCHiZrgKzs [00:20:21]). URP 6.3 has Fixed exposure only: one Post Exposure per zone, lights solved from an 18% gray card (the unaided sun 2.0 read +1.03 EV with no tonemapper; the solved 0.98 read +0.16 EV under Neutral).
2. Expect gray after baking; fix it in post under the tonemapper you chose early (Kyle Banks, 1agSNKuAfTM [00:04:58]; PYD [01:01:29]): Neutral for lookdev; contrast, saturation, Shadows Midtones Highlights pushed too far, then dialed back. ACES darkens (K3-wPnhmDi4 [00:54:50]; measured -1.4 EV): re-solve exposure under it.
3. Shadow distance before resolution (6.3 Manual; HCXCmHgV7Sk [00:15:32]): 1024 over 10 m beats 2048 over 40 m. Max Distance = farthest dynamic shadow the player sees; last cascade ends at the last caster. Stylized room with soft shadows: 1024 was the sweet spot, higher looked "too crisp" (URP e-book).
4. The asset that renders is the Quality level's, and so is its post (URPS [00:03:14], [00:30:06]). Report before editing. Unity 6 stacks the Graphics default profile, the active URP asset's profile, then scene volumes, per parameter: tier-dependent post (bloom cost, DoF, APV leak reduction) goes on each tier's asset profile, never in scene volumes, which win on every tier (measured). The template's two assets share one profile, and its Neutral tonemapper.
5. Bake what does not move; split settings by cost (Ciro Continisio, hMnetI4-dNY [00:35:51]): texel density, max size, directional mode and shadowmask cost memory, samples and bounces only bake time. Small props keep Contribute GI with Receive GI = Light Probes ([00:48:30]). A shadowmask texel holds 4 shadowed Mixed lights (6.3 Manual). Lightmaps sample 3 to 4 times cheaper than APV but take 3 times the memory (APV24 IpVuIZYFRg4 frame 00:23:43): phones keep lightmaps for static geometry.
6. The sky is the default interior light (PYD22 DlxuvvYZO4Q [00:08:17]; U25 [00:45:56]). Without baked GI and a local box probe, interiors and glossy surfaces take the sky (a smoothness-0.3 table read gray-blue until one probe existed).
7. Stylized is a production choice, not a filter (Kyle Banks [00:22:04] to [00:23:06]): one palette texture, triplanar world noise, one shared shader (7 samples over 3 textures, fine on Switch); Simple Lit on the low tier (URP e-book); Subtractive as the cheapest mixed mode (TRB). A toon shader must still sample lightmaps, shadowmask and probes or the bake vanishes; 6.3 builds custom lighting on the URP Unlit target (deltas section 5).
8. On tilers, bandwidth is the budget; prove it (6.3 Manual; U25 [00:30:26]; Ryan Keeble, bl0vKZkzNOs [00:14:30]): Forward, no depth or opaque texture unless needed (copy After Transparents), no soft shadows, MSAA 2x, 1 to 2 ms per extra full-screen pass on a Quest 2-class GPU. PassAudit reads the Render Graph's native passes: a depth copy After Opaques or the Opaque Texture splits the opaque, skybox and transparent pass (measured).

Establish first

Ask once: platforms and device tiers (compute or GLES); frame budget (mobile at 65%); one emotion sentence (KB [00:03:25]), references, stylized or realistic; time of day; what moves; interiors; custom shaders; baked-memory budget.

Decision	Default	Change when
Pipeline	URP 17.3	HDRP only for PC/console needing its exclusive features; no new features, no mobile, web or Switch 1 (rps)
PC tier	Forward+, depth texture, MSAA 4x, HDR 32-bit, 2048 shadows over 60 m in 2 cascades, soft Medium, additional shadows, Distance Shadowmask (verified on Forward+), GPU Resident Drawer when draw calls bind the CPU	device tests
Mobile tier	Forward, no depth/opaque texture (the Opaque Texture also silently drops MSAA without StoreAndResolve), MSAA 2x, render scale 0.85, 1024 shadows over 30 m, 1 cascade, soft off, no additional shadows, Shadowmask mode, bias 0.1 / 0.5 (template 1 / 1 leaked)	a wide device spread: add a low tier with copy_from (Simple Lit, no realtime shadows) [added]
Post per tier	one Grading Mode (HDR) and LUT 32 everywhere; Mobile profile: bloom Kawase or Dual (6.3), no HQ filtering, Half or Quarter, at most 5 iterations, no DoF, no SSAO, Leak Reduction Performance; PC: SSAO Downsample, Leak Reduction Quality	art direction
GI	Static geometry lightmapped with a Mixed sun (Shadowmask where a phone tier bakes static sun shadows, else Baked Indirect); APV for dynamic objects; open or changing layouts: APV only	Subtractive for low-end stylized (CC [00:41:17]); APV scenarios or sky occlusion for time of day
Probes	Adaptive Probe Volumes	Light Probe Groups without compute or when probes must move (APV24 [00:08:31])
AA	PC MSAA 4x or TAA/STP; mobile MSAA 2x or FXAA	TAA/STP never with MSAA, stacking or dynamic resolution; STP not on GLES

Workflow

Channel: ut_run.run_method(P, "AgentKit.Lighting....", args, graphics=True) for anything that renders or bakes; -nographics only for reports, tiers, volumes and audits (it cannot bake GI: Bake refuses). Plain command (L.raw_command): Unity -batchmode -nographics -quit -projectPath P -logFile <job>/unity.log -executeMethod <method> -agentJob <job>, args in <job>/args.json, result on the AGENT_RESULT line. Full code and results: references/procedures.md.

1. Report the pipeline. LightingPipeline.ReportPipeline: asset per level, renderer, features, asset profile, GRD prerequisites. GATE: every level has its own asset.
2. Tiers as data. ConfigureTiers (asset_props, renderer_props, quality_props, renderer_features, volume_profile + volume_overrides, project graphics for BRG variants and static batching), then ReportPipeline in a new process and L.tier_check. GATE: 0 mismatches, 0 errors (grading mismatch, GRD chain, APV masks without Use Rendering Layers).
3. Bookmarks and cards. AgentView_* cameras, gray cards (sun, shade, one per zone), a chrome ball indoors; CaptureStage. GATE: L.stage_review 0 errors; sheet opened.
4. Exposure, then lights. Lookdev volume (Neutral, fixed Post Exposure); sun I * 2^-EV(card); other lights by ratio; zone exposure volumes at priority 1. GATE: sunlit card within half a stop; sun-to-shade at most 4 stops [added].
5. Bake. LightingAudit.AuditLighting first (contributors, lightmap UVs, shaders that cannot show or feed the bake, Shadowmask overlap, style: "stylized" rules), then LightingBake.Bake preview, then final. Verify the Lighting Mode each tier really renders with the shadowmask range scene (P13). GATE: bake report within budget, no UV warnings, audit 0 errors.
6. Reflections and leaks. Box probes per zone (importance 2, blend 0.25 m). Leak ladder: wall thickness near probe spacing, APV Options biases (0.1 / 0.175), Rendering Layer masks (apv_layers, max 4, rebake), Probe Adjustment Volume last (APV24 [00:18:13] to [00:23:13]). GATE: mid-smooth surfaces not sky-blue indoors (blue-minus-red at or below 0).
7. Grade. Final look in the scene volume, cost fields in the tier profiles, exposure re-solved. GATE: saturation and contrast above the baked stage, clipped at most 5%, cards in band.
8. Measure. AgentProfile + ut_stat.budget_check; PassAudit per tier; PC vs Mobile captures; a development player for tier-only oddities (P12) and the PSO trace and progressive warm-up (P15). GATE: budget as a trend, device verdict after a thermal soak with a vendor GPU capture (scenario-unity-mobile); Mobile within 0.35 EV and 15% saturation of PC [added].

Numbers

Value	Relative to	Source
EV100 14 sun, 10 overcast, 8 sunset, 6 bright interior, 1 dark interior	real-world anchors; URP has no EV, use ratios	PYD [00:17:37]
51 vs 102 texels/m	2048 over 40 m vs 1024 over 10 m	6.3 Manual
spot 1, point 6, max 16 maps	additional-light shadow atlas	6.3 Manual
256 samples; 4 lights per texel	denoiser floor; shadowmask channels	6.3 Manual
0.2 background, 1 play area	Scale In Lightmap	CC [00:17:30]
0.20 vs 0.67 to 0.83 ms; 42.7 vs 14.5 MB	lightmaps vs APV sampling (PS4 1080p); memory	APV24
1,723 / 9 / 0 / 0 px	sun leak on Mobile: template bias, 0.1 / 0.5, Shadowmask, Subtractive	observed
1.33 vs 0.67 MB	interior lightmap set, Shadowmask vs Subtractive (512 x 512)	observed
about 2,450 to 22 to 38 draw calls; main thread / 3	900 cubes, GRD off vs on, Editor Play mode	observed
6 / 8 / 7 native passes	Mobile: depth off / copy After Opaques / After Transparents	observed
14 PSOs, warmed in 5 frames	outdoor scene, Metal development player, 3 per frame	observed
1 to 2 ms per full-screen pass	Quest 2-class tiler	RK [00:14:30]

Quality gates

• Measurable: fresh-process tier readback 0 mismatches; ReportPipeline and AuditLighting 0 errors; bake within budget; cards in band under the final tonemapper; clipped at most 5%, crushed at most 30%; chrome ball not sky-blue indoors; far shadows present where the tier promises them (P13); opaque, skybox and transparent in one native pass on phones; budget_check as a trend.
• Visual: the stage sheet after every stage; crisp shadows without acne or seams; no leaks at joints; no sky tint; the emotion sentence describes the frame; Mobile reads as the same image (references/critique.md).

Common mistakes

Mistake	What it looks like	Fix
Lights by feel, exposure to rescue them	card +0.9 EV, interiors never balance	gray card, fixed exposure, solve lights
Resolution raised instead of distance lowered	4096 still soft near the camera	Max Distance to the last visible dynamic shadow
CPU lightmapper or -nographics bake on this Mac	refused, or nothing baked	GPU lightmapper, graphics=True; there is no CPU fallback on Apple Silicon
Small props with Contribute GI off	no baked contact shadow; realtime shadows under Shadowmask	Contribute GI on, Receive GI = Light Probes
More than 4 shadowed Mixed lights on one surface	extra lights frozen as fully baked	light.shadowmask_overlap; make extras Baked or Realtime
Toon or custom shader on lightmapped geometry	baked light and shadows gone	light.shader_no_lightmap / no_shadowmask / no_meta; scenario-unity-shaders adds the variants
Per-tier bloom set in the scene volume	Mobile keeps PC bloom cost	cost fields only on tier profiles
One tier LDR, one HDR	tiers drift apart after grading	one Grading Mode (HDR), one LUT size
APV masks without Use Rendering Layers	masks ignored on that tier	m_SupportsLightLayers on every asset
No local reflection probe indoors	chrome and wood reflect sky	box probe per zone, importance 2
Depth Priming on with custom shaders	objects vanish once MSAA is off	DepthOnly and DepthNormals passes, or priming Disabled
Template Mobile bias 1 / 1 on thin walls	sawtooth sun along the wall-ceiling joint	bias 0.1 / 0.5; bake the sun (Shadowmask or Subtractive)
GRD on without its chain	nothing batches	BRG Variants Keep All, SRP Batcher, Forward+, static batching off

Handoffs

• Receives: blockouts from scenario-unity-world-building (static flags, lightmap UVs checked here); shaders from scenario-unity-shaders (DepthOnly, DepthNormals, Meta passes; LIGHTMAP_ON and SHADOWS_SHADOWMASK variants when lightmapped); budgets from scenario-unity-performance and scenario-unity-mobile.
• Delivers: tiers (JSON spec + readback), tier profiles, bakes with their report, stage sheets, pass audits. To scenario-unity-shaders: the stylized shading brief (palette, shared shader, ramp on the URP Unlit target that samples lightmaps, shadowmask and APV; Simple Lit variants for the low tier) and Render Graph passes. To scenario-unity-performance: GRD bench CSVs and GPU questions. To scenario-unity-mobile: the Mobile tier, PSO traces per graphics API, the thermal soak and vendor GPU captures (Xcode, Android GPU Inspector, Arm Performance Studio, Snapdragon Profiler). To scenario-unity-pipeline-automation: raw command lines for CI. 2D Renderer lights: scenario-unity-2d.

Unity 6.3 notes

• Render Graph only; _CLUSTER_LIGHT_LOOP replaces _FORWARD_PLUS. The Render Graph Viewer can attach to a device player; PassAudit reads internals by reflection (available: false if they move).
• URP exposure Fixed only, no physical light units, SSR or volumetrics; HDRP in maintenance; Built-in deprecated in 6.5.
• The feature table says Distance Shadowmask is "Forward rendering path only" (it says the same of MSAA); on 6000.3.21f1 it rendered on Forward+ exactly as on Forward (P13).
• GPU lightmapper only on Apple Silicon; CPU lightmapper deprecated in 6.6; OptiX deprecated in 6.5.
• Bloom filter: Gaussian, Dual, Kawase (new in 6.3). GraphicsStateCollection is experimental. APVLeakReductionMode.ValidityBased is an obsolete alias that ToString() prints for Performance.
• Dynamic batching deprecated in 6.5. STP forces TAA, needs compute, never GLES.

References

• references/procedures.md: procedures P1 to P16 as copyable code, each with its live test and result.
• references/critique.md: the rubric for captures, cards, bakes, reflections, grade, tiers, bandwidth.
• references/expert-notes.md: principles by expert, the stylized recipe, disagreements, observations.
• references/gui-paths.md: windows and menus for the same procedures.
• references/sources.md: sources, credentials, timestamps, baseline gaps, revision history.
• scripts/ut_lighting.py, scripts/AgentKit/Lighting/*.cs, scripts/Runtime/*.cs (player helpers: verification builds only).