Blender light camera

blender-light-camera — half of perceived quality is presentation

The same mesh can look like a student exercise or a product shot; the difference is light, camera, and color management. Weak agents render with the default point lamp, a 50mm camera jammed too close, and the Standard view transform — three defaults that each cost more quality than a day of modeling. This skill replaces them.

Color management FIRST — before judging any lit render

• View transform: AgX (Blender 4.0+) or Filmic (2.8x–3.x). Standard clips highlights and flattens color — beauty renders judged under it are judged wrong. Probe the enum if unsure: assigning an invalid name raises, and the error lists nothing — read bpy.types.ColorManagedViewSettings.bl_rna.properties['view_transform'] or just try AgX→Filmic in order.

import bpy
vs = bpy.context.scene.view_settings
vs.view_transform = 'AgX' if bpy.app.version >= (4, 0, 0) else 'Filmic'
vs.look = 'AgX - Base Contrast' if bpy.app.version >= (4, 0, 0) else 'Medium High Contrast'
vs.exposure = 0.0   # iterate: render, look, adjust in ±0.5 steps

• Exposure is iterated like everything else: render → histogram-by-eye (highlights not clipped, shadows not crushed) → adjust → re-render. Two rounds usually suffice.

The scripted three-point rig (starting values — then iterate by render)

import bpy
from mathutils import Vector

def rig_three_point(center, size):
    for old in [o for o in bpy.data.objects if o.name.startswith("Rig_")]:
        bpy.data.objects.remove(old, do_unlink=True)
    d = max(size)
    spots = {"Key":  (( 1.2, -1.0, 1.4), 4.0),   # position (×d), relative power
             "Fill": ((-1.4, -1.2, 0.6), 1.0),
             "Rim":  ((-0.5,  1.4, 1.1), 2.5)}
    for name, (off, power) in spots.items():
        ld = bpy.data.lights.new(f"Rig_{name}", 'AREA')
        ld.size = d * (1.5 if name == "Fill" else 0.8)
        ld.energy = 300 * power * d * d          # scales with subject size; iterate
        lo = bpy.data.objects.new(f"Rig_{name}", ld)
        bpy.context.scene.collection.objects.link(lo)
        lo.location = center + Vector(off) * d
        lo.rotation_euler = (center - lo.location).to_track_quat('-Z', 'Y').to_euler()

• Ratios: key : fill ≈ 3–4 : 1 (fill exists to open shadows, not to flatten); rim separates the silhouette from the background — check the 3/4 beauty for a readable edge light.
• Big soft lights (area size ≈ subject size) flatter forms; small hard lights dramatize and expose surface flaws — choose per art direction, say which.
• Ground: a plane ~10× the subject named Rig_Ground (the Rig_ prefix keeps rig objects out of probe framing — see blender-verify), neutral gray albedo ~0.35–0.5, roughness ~0.6 — grounds the object with a real contact shadow. Floating objects on void backgrounds read as unfinished.
• HDRI alternative: if an .hdr/.exr exists in the project, World → Environment Texture beats hand rigs for realism (keep the rim light). Don't fake an HDRI from memory of one.
• Motivated light for scenes: pick the story source (window, lamp) as key; blender-scene owns scene-level composition.

Camera discipline

• Focal length: 50–85 mm for products and subjects (default 50 mm is fine — the crime is distance, not lens). Wide (<35 mm) only for interiors/environments, and then watch verticals.
• Distance so the subject fills ~65–75% of frame height; headroom above, more space in front of the subject's "facing" direction than behind.
• Height: eye-level for neutral catalog reads, slightly below subject mid-height for heroic presence. Straight-on axis shots read as diagrams — the money shot is the 3/4.
• Composition: rule of thirds for off-center subjects; for symmetric hero shots, deliberate dead-center symmetry — pick one, don't drift between.
• Depth of field: OFF by default; a subtle f/2.8–f/4 for hero close-ups only, focus on the identity feature.

Engine per purpose

Purpose	Engine
Form/silhouette probes, wireframe	Workbench (fast, headless-safe, honest about shading)
Iteration beauty during material/light work	EEVEE (BLENDER_EEVEE_NEXT from 4.2, BLENDER_EEVEE before — version-gate it); enable AO
Final hero beauty + turntable	Cycles, 128–256 samples, denoise on (CPU works headless; GPU if available)

scene = bpy.context.scene
if hero_final:
    scene.render.engine = 'CYCLES'
    scene.cycles.samples = 256
    scene.cycles.use_denoising = True
else:
    scene.render.engine = 'BLENDER_EEVEE_NEXT' if bpy.app.version >= (4, 2, 0) else 'BLENDER_EEVEE'

Materials that need transmission/subsurface silently degrade in the wrong engine — final material judgment happens in the final engine. Caveat: EEVEE needs a GPU/display even under --background; on a truly headless server, form-probe with Workbench and do all beauty work in Cycles (CPU renders fine, just slower).

Presentation set (what "done" renders are, per tier)

• production: lit 3/4 beauty at ≥1280px + the 4-view probe set.
• hero: the above + 8-step turntable + 2-3 close-ups of identity features/detail zones (snippets in blender-verify), all under the same rig and color management — a presentation set with inconsistent lighting reads as accident.

Then look at every image and score "lighting & presentation" on the rubric with named evidence (highlight clipping? silhouette separated from background? contact shadow present? composition rule stated?).