Blender materials

blender-materials — realism is mostly roughness, not color

Untrained material work has one smell: every surface is a single uniform value, so everything reads as fresh-from-the-factory plastic. Real surfaces vary — mostly in roughness. This skill turns art-direction words ("worn brass", "matte lacquer") into node-level decisions, and keeps material judgment separate from form judgment.

Clay first, beauty second

Form is judged on clay renders (uniform gray override — snippet in blender-verify); materials are judged on beauty renders under the real light rig. Never debug both at once: a render that looks wrong is first split into "form problem or surface problem?" by comparing the clay and beauty probes of the same view.

Principled BSDF values that read as real

Rule	Numbers
Albedo has a floor and a ceiling	nothing below ~0.03 (charcoal) or above ~0.85 (fresh snow); "black" plastic ≈ 0.03–0.05, "white" paint ≈ 0.7–0.8. Pure 0/1 kills light transport
Metallic is binary	0 or 1; the 0.5 middle exists only for dusty/painted metal masks. Metal color lives in Base Color (gold ≈ (1.0, 0.75, 0.35))
Roughness is the king	judge every material first by its roughness story; see variation below
IOR	leave ~1.45 for dielectrics; glass 1.52, water 1.33. Don't sculpt specular response with the specular slider — that's IOR's job
Saturation restraint	realistic albedo saturation is lower than instinct (most real objects < 0.6); full-sat colors read as toy plastic
Subsurface / transmission	only with a reason (skin, wax, food, glass) and only in Cycles or configured EEVEE — verify in a render, these settings silently no-op in wrong engines

Roughness variation — the single highest-yield realism move

Uniform roughness = CG. Recipe (values are starting points):

• Base roughness by material (polished metal 0.15, brushed 0.35, wood satin 0.4, matte paint 0.6, rubber 0.8).
• Add a Noise Texture (scale ≈ 4–8 relative to object size) → ColorRamp squeezed narrow (e.g. stops at 0.45/0.55) → into Roughness, remapped to ±0.05–0.15 around the base. Subtle: visible as broken-up highlights, not as dirt.
• Wear at contact points (art direction permitting): lighter/rougher (or darker/smoother — decide per material story) on edges people touch. Cheap proxy without UVs: a second noise masked by Geometry→Pointiness through a ColorRamp — Cycles only; in EEVEE, bevel-edge highlights carry most of the effect alone.

Defensive node scripting (input names drift across versions)

import bpy

def make_pbr(name, base=(0.5, 0.5, 0.5, 1.0), rough=0.5, metal=0.0):
    mat = bpy.data.materials.get(name) or bpy.data.materials.new(name)
    mat.use_nodes = True
    bsdf = next(n for n in mat.node_tree.nodes if n.type == 'BSDF_PRINCIPLED')
    def set_in(sock, val):
        s = bsdf.inputs.get(sock)
        if s is not None:
            s.default_value = val
        else:
            print(f"PROBE material_input_missing {name}:{sock} -> {list(bsdf.inputs.keys())}")
    set_in("Base Color", base)
    set_in("Roughness", rough)
    set_in("Metallic", metal)
    return mat

def add_roughness_variation(mat, base_rough=0.4, amount=0.1, scale=6.0):
    nt = mat.node_tree
    bsdf = next(n for n in nt.nodes if n.type == 'BSDF_PRINCIPLED')
    noise = nt.nodes.new('ShaderNodeTexNoise')
    noise.inputs["Scale"].default_value = scale
    ramp = nt.nodes.new('ShaderNodeValToRGB')
    ramp.color_ramp.elements[0].position = 0.45
    ramp.color_ramp.elements[1].position = 0.55
    ramp.color_ramp.elements[0].color = (base_rough - amount,) * 3 + (1.0,)
    ramp.color_ramp.elements[1].color = (base_rough + amount,) * 3 + (1.0,)
    nt.links.new(noise.outputs["Fac"], ramp.inputs["Fac"])
    nt.links.new(ramp.outputs["Color"], bsdf.inputs["Roughness"])
    return mat

inputs.get() + the printed fallback is the pattern: name drift becomes a visible probe line, not a silent no-op or a crash. Missing-input prints get resolved by reading the printed key list, not by guessing again.

Palette discipline

• 60-30-10: one dominant material family, one secondary, one accent. A part list where every part got its own loud color reads as a toy.
• Separate parts by VALUE (light/dark) first, hue second — value contrast survives lighting changes and grayscale probes; hue contrast doesn't.
• The color-block materials from the build phase are placeholders — replace them by the part list, don't decorate them.

UVs — only when something needs them

Procedural (noise/pointiness-driven) needs no UVs. You need UVs when: image textures are supplied, text/logos must land exactly, or the target is a game engine (then: Smart UV Project is acceptable for props — probe for stretching in a checker render; hand-seamed UVs are beyond scripted fidelity for complex organic shapes, say so). For tileables without exact placement, box mapping (Texture Coordinate→Object + Mapping node) avoids UVs entirely.

Verify hooks

Every material lands in the next beauty probe: check highlight breakup (roughness variation visible?), value separation between parts, metals actually reading as metal (dark environments make metals black — that's a lighting note for blender-light-camera, not a material bug). Rubric axis "material realism" scores against these, with the render as evidence.