Lighting and rendering in Maya 2027 (Arnold)

Expert lighting is a story decision made measurable: know where the eye must land, shape every form with one dominant light, structure values back to front, prove it with light groups and numbers before the eye signs off. Rendering is engineering: find the noise in the AOVs before touching a sampler, deliver scene-linear EXRs whose passes rebuild the beauty, judge through a tone-mapping view. 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-maya-expert (execution channel, review loop, 2027 version traps).

Status (2026-09-24): scripts/mx_light.py is not yet run in Maya: 138 offline checks passed (python3 + numpy); the Maya layer ran only against a fake maya.cmds. Once Maya is installed: tests/code/maya-lighting-rendering/run_all.sh, then fix names from probe_lighting.json.

Stance (the expert delta)

• Story and target first; brighten by darkening. "What is the story I am trying to tell?" before any button; the target may be a hand or a door, not the face by default (KT [00:16:50] [00:15:49]; TZ-C [00:10:51]). To brighten the subject, darken its surround; pushing the subject clips it (KT [00:30:31]).
• One light stands above the others. "Too many light sources at the same intensity" is the common failure (BR-B ch6 Balance); a rim hotter than the key fails too: "small and hot" is radiance, not more energy (JHILL [00:13:47]). Arvid keeps the key strongest (ARV-L [00:06:51]); for Brejon it is the main shaping light (BR [00:23:36]). Decider: pick the key by the shape it gives, then measure ONE dominant group on the subject mask.
• Key upstage, measured from the face axis. Light the far side of the face (BR-B ch8.5); measured from the lens axis, a turned face gets a frontal key [added test]. No light on the camera side (BR [00:13:46]); the wrap is front-ish, lower, more saturated, about twice the key's size (BR [00:26:54] [00:27:59]); never an ambient light.
• Values back to front, counterchange behind every subject. Far blacks lifted and flat, foreground darkest and most contrasty (TZ-C [00:08:38]); light behind shadow and shadow behind light, never dark on dark (BR [00:38:56] [00:40:02]); vignette the lighting with blockers (BR-B ch6 Vignetting).
• Reference first; on polished products the reflections are the product. Real photographs of the same subject and situation, their light placement rebuilt before judging any material (ARV-C [00:02:44] [00:36:21]); Brejon lost two days without one (BR [00:08:13]). Each softbox sits on the reflected view ray of the plane it defines (reflection_placement) [added], black flags give dark bands; soft means large (JHILL [00:08:26]).
• Find the noisy AOV, then raise that sampler only. The wrong rays cost time and leave the noise (SMP § Removing Noise); rays per pixel = AA² x samples², so raising AA means lowering the rest (ARV-T [00:07:12]). Maya 2027 uses Global Light Sampling: one count, 4 or fewer on CPU; per-light samples only matter on dome, distant and local-mode lights (SMP § Limitations). Emission is the noisiest light (LGT § Mesh Light vs Emission).
• 90% of the look in the beauty; grades go back into the lights. Shaping cannot be faked in comp (KT [00:45:39]); light groups answer small notes (SARK [00:11:28]); "I ALWAYS copy back the values" (BR-B ch8 Final result) as a multiplier on the light color (BR [01:03:30]).
• Judge through the ACES view, deliver scene-linear. Without tone mapping you under-light into "a complicated rig of 50 lights" (BRJ-CM Ch.1); output transform off for EXR (CM); under ACEScg a linear HDRI is linear Rec.709, not Raw (CM § ACES Workflow).

Establish first

Input	Changes	Default when silent
Deliverable: still or sequence, size, range	sampling, denoiser, budget	still; block at 320 px, preview at half size
Story and focal target	every gate	product: the dial or logo; character: the speaking face
References	light placement, values	photos of the same subject and situation in brief["references"]; the critique warns without
Machine	device, denoiser	macOS: CPU only, no Arnold GPU or OptiX; OIDN
Mood, art direction	face lit/shadow ratio, symmetry	"drama", 2 stops [added MOOD_RATIO_STOPS]; natural, asymmetric (BR-B ch6 Planet 51)
Continuity contract	allowed shot tweaks, one master per camera axis	conservative: exposure, transform, visibility (BR [00:37:17])
Color pipeline	rendering space, view, delivery	Maya default: ACEScg, ACES view, EXR ACEScg half
Comp needs	AOV preset, light groups (16 max)	groups on RGBA, Z, subject mask; comp preset for a compositor
Render budget	settings ladder	measured on a preview, extrapolated before a final
DOF and motion blur	3D or comp; what must be sharp	3D for hair, glass, overlapping depth (BR [00:54:45]); bokeh needs small bright sources (BR [00:55:51])

Workflow

Headless through scenario-maya-expert's mx_run.py --plugins mtoa (import mx_light as L); frame analysis runs in plain python3. Arnold RenderView is GUI only: iterate with batch renders you open.

1. Intake. mx_validate.validate(profile="shot") on shots; scenario-maya-lookdev's handoff_report for assets; L.cm_audit_scene(); write the brief (L.default_brief, procedures P1) with its references; read in each where the key, rims and dark bands sit. GATE: no validation fail, color audit pass, story, target and references written.
2. Plan and block. L.plan_product_rig(name, bb_min, bb_max, cam_pos) or L.plan_character_rig(char, head_c, head_size, cam_pos, face_forward); L.rig_plan_checks(plan); L.apply_rig(plan). Brejon's order: practicals, key, wrap, rim, top; atmosphere from the first render (BR [00:19:46] [00:33:29]). Render on gray (with L.grey_shading(0.18), preset block), open L.light_group_sheet(exr, png). Aperture from L.dof_plan(near, far, focal) on L.view_depth_range of what must be sharp. GATE: key upstage, wrap off the lens axis, 16 groups or fewer; each light solo does its job (BR [00:25:14]); DOF plan holdable.
3. Shape and balance by measurement. r = L.render_shot(out, cam, masks={"subject": [root]}), rep = L.analyze_frame(png=r["png"], exr=r["exr"], brief=brief), L.write_report(rep, out). Work the findings top-down (reference, eye path, values, separation, balance, face, color). GATE: no fail; every warn fixed or kept with the reason; annotated.png, squint.png and the solo sheet judged with references/critique.md.
4. Master to shots (sequences). A lights-only master per camera axis, referenced into each shot, rendered out of the box first, widest shot first; tweaks only within the contract, logged (procedures P13; BR-B ch8). GATE: L.compare_frames against master and neighbors; signature features (a character's rim) in every shot.
5. Noise and budget. pair = L.render_seed_pair(out, cam) (denoiser off), L.noise_pair(pair["a"], pair["b"]), L.next_step(...), or L.noise_loop(render_pair, settings). Fireflies go to L.FIREFLY_PLAYBOOK, not to samples. OIDN first in the chain (L.set_denoiser(first=True)), noice for sequences; OIDN forces a box filter and denoises full frames (AOV § OIDN): cut detail crops from a full frame, render the detail_loss reference with a box filter. L.extrapolate_time, L.sequence_budget, L.TIME_SINKS before cutting samples. GATE: beauty noise at the brief's target under the chosen denoiser; L.detail_loss about 0.8 or more on hero detail; time fits.
6. Passes and comp handoff. L.set_light_groups(one_dict), L.setup_aovs("comp"), L.add_mask_aov, L.setup_exr_driver(); L.check_sums, L.group_contract, L.comp_recipe. Comp notes return through L.grade_to_light. GATE: groups plus default rebuild the beauty within 1%, no empty requested pass, EXR single-part, zip, half color, float Z/P/N.
7. Final and delivery. Render Setup layers as a template (L.make_layer, L.export_render_setup), L.render_cmd(...) with -rst, or L.render_shot per frame; .tx pre-baked, auto-TX off, use-existing on (ARV-T [00:14:44]); L.licence_check_cmd(), then L.render_env(final=True) so an unlicensed batch aborts instead of watermarking; re-time one full frame; check exit code and files; critique; new version. GATE: frames on disk, report written, not-verified list stated.

Numbers

Item	Value	Relative to, source
Rays per pixel	AA² x samples²: AA 3 diffuse 2 = 36; AA 6 specular 6 = 1296	SMP § Camera (AA)
Global light samples	4 on CPU; 1 or 0 when a dome or distant light dominates	SMP § Best setting
Adaptive	AA 3 to 4, Max AA 20, threshold 0.015 [verify 2027 default]; off in previews	SMP
Motion blur, DOF	AA about 12, up to 20 for strong DOF, secondaries 1	ARV-T [00:20:29]
Ray depth	diffuse 1 to 2 products and exteriors, 3 to 4 interiors; specular 2 unless glass or inter-reflecting metal	ARV-T [00:05:59]; SARK [00:08:27]
Skydome resolution	1k for lighting only; the HDRI width (cap 8k) when chrome or glass reflects it	ARV-T [00:07:54]; LGT § Resolution
DOF	aiApertureSize = radius (world units) = focal/(2N); sharp depth about c d / radius, c = frame width/1200 [added]; same framing, same depth at any focal	LGT § Aperture Size; L.dof_plan
Clamps	indirect 10 or more; pixel clamp 30 to 50	BRJ-CM Ch.1 Range
Light AOVs	16 max; Brejon used 9 for 30 lights	AOV; BR [01:12:25]
noice	strength 0.45 (0.2 texture, 0.8 GI and SSS), -ef up to 2, no -t on macOS	AOV § Arnold Denoiser
ACES view	scene 1.0 shows about 0.81; a bit more than 16 reaches white	MLC [00:02:13]
Wrap	2x the key's size, lower; 30 deg or more off the lens axis [added threshold]	BR [00:27:59] [00:13:46]
Starting exposure	E = log2(pi d² target), normalized area light at distance d; 15 at 1 m in cm	[added] from LGT § Exposure, § Normalize; Arvid's 15 to 20 (ARV-L [00:03:58])

Quality gates

• Code: rig_plan_checks pass; scene_checks no fail (ambient, Samples 0, HDRI Raw under ACEScg, dome resolution, group typos, key upstage); render_settings_report answered (denoiser, TX, device); analyze_frame verdict passes; check_sums ok; cm_audit pass; the noise loop stopped for a stated reason. L.THRESHOLDS are [added]: calibrate on an approved frame.
• Visual: the beauty at full size and as squint.png, beside its reference; annotated.png (saliency peak, clipped red, crushed blue); the solo sheet; the -5 to +5 stop bracket; sequences cut with neighbors. Judge with references/critique.md, in order.

Common mistakes

Mistake	Looks like	Fix
Key measured from the camera	frontal, flat key on a turned face	plan_character_rig measures from the face axis
Fill from the lens axis, or an ambient light	flat faces, no core shadow	wrap front-ish off axis; delete ambient (unsupported in Arnold)
Rim equal to or above the key	two competing sources	rim at least 0.5 stop under the key; rim_equals_key
Raising AA or every sampler for noise	time explodes, noise stays	noise_pair, then one sampler; AA for alpha, blur, reflected diffuse
Per-light samples raised on area lights	no effect under GLS	global light samples; per-light only for dome, distant or local mode
Emissive geometry as the light	noise that samples cannot clean	area or mesh light (LGT)
Light-group typo	empty pass, no error	one dict; group_contract, check_sums
A light group for bounce cards or an emissive plane	empty RGBA_cards	groups live on lights; a card lands in its light's group; emission is O (emission AOV, LPE C.*O)
Aperture from habit ("f/11 holds a product")	half the product soft	L.dof_plan; camera nearer perpendicular to what must be sharp
Dome at 1k while chrome reflects a 4k HDRI	soft reflections	resolution at the HDRI width
HDRI or EXR left on the Raw rule under ACEScg	oversaturated, wrong lighting	linear Rec.709 input space
Old light values copied into 2027	renders look dark, rig grows	re-judge exposure through the ACES view

Handoffs

• Receives from scenario-maya-lookdev: <asset>_lookdev_v###.ma, a passing handoff_report() JSON, the turntable sheet, sampling notes (SSS, transmission depth, detail to protect from the denoiser). From scenario-maya-animation: a shot version, rigs referenced, validate(profile="shot") clean, the approved playblast. From scenario-maya-fx and scenario-maya-groom: versioned caches with frame ranges.
• Delivers to comp or the client: EXRs <shot>_<layer>_v###.####.exr (ACEScg scene-linear, single-part, zip, half color, float Z/P/N, light groups, masks); display PNGs; critique.md/json, annotated.png, solo and bracket sheets; comp_recipe JSON; the render log (seconds per frame, rays per pixel); the not-verified list. Back to scenario-maya-lookdev when a fix must change GI or reflections, or repeats across shots (BRJ-CM Ch.9).

Maya 2027 notes

• MtoA 5.6.x (Arnold 7.5, license ARNOL_2027); Arnold RenderView, since Maya IPR does not work with MtoA 5+.
• Global Light Sampling by default, per-light sampling mode in 5.6.0; per-group indirect AOVs exclude emission (Arnold 7.3), so sums need RGBA_default and emission.
• macOS: settings_lint fails a GPU device; OIDN in new scenes; no noice -t.
• Batch renders abort on license failure by default (Arnold 7.3, abort_on_license_fail); mx_run sets WN25's override ARNOLD_FORCE_ABORT_ON_LICENSE_FAIL=0 so tests watermark [verify value]. Render -r arnold, never maya -render.
• OCIO v2: ACEScg, ACES view; input sRGB Encoded Rec.709 (sRGB) since 2026.2; no ACES2065-1 rendering space; .tx, .hdr, .exr rules default to Raw; query names.
• Render Setup and legacy layers are exclusive per session; setups travel as JSON templates.

References

• references/expert-notes.md: depth per expert with timestamps, disagreements and deciders. Load when a decision is not covered above.
• references/procedures.md: blocks P0 to P13, brief to master and shot. Load before writing lighting code.
• references/critique.md: frame critique mapped to code checks, rubrics, severity. Load at every gate.
• references/gui-paths.md: menus and editors for a computer-use agent or a human.
• references/sources.md: sources, credentials, URLs, timestamps.
• scripts/mx_light.py: the toolkit; its docstring lists every call and where it runs.