Scenario Maya Lighting Rendering

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

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.