specular_color tinting of internal reflections is inconsistent in the reference MaterialX graph (TIR untinted, partial internal reflection tinted)

[portsmouth]

Summary

The reference MaterialX graph for open_pbr_surface does not distinguish between external and internal reflection when applying specular_color, and as a side effect it tints the two kinds of internal reflection inconsistently:

• partial internal reflection (below the critical angle) is tinted by specular_color;
• total internal reflection (TIR) (above the critical angle) is not tinted by specular_color.

This came up in the working-group discussion around specular_color / back-face behavior; filing it so it can be resolved in the spec and reflected in the reference graph (and noted for the ASWF/OpenUSD normative shading-model docs).

What the graph does

The dielectric base is built from two independent single-mode lobes:

Node	scatter_mode	tint
dielectric_reflection (+ _tf thin-film variant)	"R"	specular_color
dielectric_transmission	"T"	if_transmission_tint (white; becomes transmission_color only when transmission_depth == 0)

dielectric_base = layer(
    top  = dielectric_reflection_tf_mix   # the R lobe, tinted by specular_color
    base = mix(dielectric_transmission(+ volume), opaque_base, transmission_weight)
)

specular_color is wired straight into tint of a single dielectric_bsdf with scatter_mode="R". That node reflects on whichever side the interface is hit (the dielectric_bsdf decides entering-vs-exiting internally from the ior), and tint multiplies the whole lobe. There is no backfacing node, geompropvalue facing test, or any view-direction conditional in the graph — so the tint applies uniformly to reflections from both hemispheres.

Verified in both:

• the OpenPBR reference graph (reference/open_pbr_surface.mtlx), and
• the shipped MaterialX library (libraries/bxdf/open_pbr_surface.mtlx, MaterialX 1.39.5).

The inconsistency

Because TIR is modeled inside the transmission lobe (scatter_mode="T") rather than the reflection lobe:

• partial internal reflection → produced by the R lobe → tinted by specular_color;
• TIR → produced by the T lobe → never tinted by specular_color.

So whether an internal reflection is tinted depends on whether the ray is above or below the critical angle — which was never intended to be an observable distinction of specular_color. There is effectively no single, well-defined behavior for "internal reflection."

Spec ambiguity

The graph matches neither interpretation cleanly:

• Literal current spec — specular_color tints the BRDF (light reflected back into the same hemisphere, from upper or lower incidence), transmission unaffected. Under this reading TIR reflections should also be tinted, but they are not (they live in the T lobe).
• "External-only" intent (raised in the working group) — specular_color should tint only the external reflection. Under this reading partial internal reflections should not be tinted, but they are (the R lobe).

Root cause / MaterialX limitation

With the current MaterialX node set, a single dielectric_bsdf(scatter_mode="R") does not expose entering-vs-exiting, so the tint cannot be gated on facing. Cleanly implementing an external-only (or a consistent internal) behavior would require either additional MaterialX functionality or a graph restructuring.

Proposed next steps

1. Decide the intended semantics: does specular_color tint (a) external reflection only, (b) all reflections including internal (both partial and TIR), or (c) something else?
2. Clarify the spec text accordingly (the current wording is being read inconsistently).
3. Once decided, correct the reference MaterialX graph — and, if it needs facing/entering-exiting information that MaterialX cannot currently express, raise the corresponding MaterialX feature request.
4. Note the current graph↔spec difference for the ASWF/OpenUSD normative shading-model documentation.

[portsmouth]

Fixing the MaterialX side — does it need a dielectric_bsdf change?

Short answer: yes, a clean fix requires changing dielectric_bsdf. From the MaterialX PBR spec, two facts settle what the graph can express:

1. tint scales reflection and transmission together — "The tint t scales the node's reflection and transmission lobes directly, multiplying the microfacet BRDF and BTDF as t·f_r and t·f_t. It does not modify the Fresnel reflectance F itself." So a single dielectric_bsdf has no way to tint its reflection independently of its transmission.
2. TIR lives inside the transmission mode — "total internal reflection may occur even when transmission modes are selected." So the TIR-reflected energy is produced by the scatter_mode="T" lobe and is scaled by that lobe's tint (in our graph, the transmission tint), never by specular_color.

This is precisely the inconsistency reported above: the explicit scatter_mode="R" lobe carries specular_color, so ordinary reflection (entering, and partial exiting/internal) is tinted — but TIR, being inside the T lobe, is not. There is no input on dielectric_bsdf to tint reflection separately from transmission, nor to distinguish entering vs exiting.

Recommended change

Give dielectric_bsdf (and, for consistency, generalized_schlick_bsdf) separate reflection and transmission tints — e.g. keep tint as the reflection tint and add a transmission_tint (default white), with the explicit definition that the reflection tint scales all reflected radiance, including the TIR-reflected component, while the transmission tint scales only transmitted radiance.

OpenPBR then maps cleanly onto a single dielectric interface (scatter_mode="RT"), so there is exactly one reflection lobe (TIR included), tinted consistently:

• reflection tint ← specular_color (applies to all reflection, incl. TIR);
• transmission tint ← white, or transmission_color when transmission_depth = 0.

This matches the OpenPBR spec's own split — specular_color tints "the entire macroscopic dielectric BRDF (for both entering and exiting rays), while the BTDF is unaffected" — and removes the TIR special case entirely.

This depends on the semantics decision above

• If the intended semantics is "tint all reflection incl. TIR, leave transmission untinted" (what the current spec text says), the reflection/transmission-tint split solves it exactly.
• If the intended semantics is "external reflection only" (as suggested in the meeting), the split is not sufficient — you would additionally need entering-vs-exiting (side) information, which dielectric_bsdf also does not expose and which is awkward to supply reliably at the graph level for a closure. That would be a larger, messier dielectric_bsdf change. This is a further argument for adopting the "tint the whole BRDF" semantics, which the spec already states and which the tint-split cleanly implements.

Interim option (no MaterialX change)

If we don't want to block on a MaterialX release, the honest interim is to document the TIR-untinted behavior as a known graph↔spec difference for the ASWF/OpenUSD normative docs. TIR-reflection only occurs at grazing internal angles, so the visible cost of leaving it untinted is limited. Either way, we should open a companion MaterialX issue proposing the reflection/transmission-tint split.

[jstone-lucasfilm]

Thanks for starting this GitHub Issue, @portsmouth, and I agree that we should open a companion MaterialX issue proposing the reflection/transmission tint split.

In the spirit of your interim option above, I've proposed new language in the latest PR for shading model documentation in MaterialX, and I'd be interested in your thoughts:

AcademySoftwareFoundation/MaterialX@a057a3a

[portsmouth]

The intended specular_color reflection semantics changed between 1.1 and 1.2 (deliberately)

Important context that reframes this issue: the intended behaviour changed between 1.1 and 1.2.

1.1 (dev_1.1): specular_color "modulates the Fresnel factor of f_dielectric, but only for the initial reflection of light incident from above, while the light transmitted from above or below (or reflected from below) is assumed to be unaffected … i.e. the reflection due to scattering from the internal medium, or the reflection from below, or the transmission from above or below." → external reflection only.

1.2 (dev_1.2): specular_color "tints … the entire macroscopic dielectric BRDF (for both entering and exiting rays), while the BTDF is unaffected." → all reflection (both hemispheres), transmission untouched.

This was a deliberate change, made in #247, per the author's comment there:

We didn't state it clearly before, but now I made it clear that specular_color functions as a (non-physical) overall tint multiplier of the total macroscopic BRDF (for both entering and exiting rays), leaving the BTDF untouched, thus removing energy. You could also choose for the tint to only apply to entering rays, but it seems unnecessary to me.

So the intended 1.2 behaviour is settled: tint the whole dielectric BRDF (entering + exiting reflection), leave the BTDF alone.

Consequences for this issue

• The reference graph's scatter_mode="R" lobe already tints reflection on both hemispheres, so it matches the 1.2 intent for ordinary reflection. The remaining defect is precisely the one reported above: TIR is emitted by the T lobe and so is not tinted by specular_color, whereas under 1.2 it should be (it is exiting-ray reflection, part of the BRDF).
• The MaterialX fix proposed above (separate reflection/transmission tints, with the reflection tint covering the TIR component) is therefore the correct direction for the 1.2 semantics. The facing/entering alternative would only be needed for the 1.1 external-only reading, which 1.2 has deliberately moved away from — so it can be set aside.

Two follow-ups worth tracking

• This is an (undocumented) look change between 1.1 and 1.2. For a transmissive or internally-reflecting dielectric with a non-white specular_color, 1.2 now tints internal reflections that 1.1 left untinted. It is not mentioned in the CHANGELOG (the specular_weight reinterpretation (as decoupled IOR) #247 entry only covers the specular_weight refraction-direction decoupling), yet it is a genuine appearance difference — worth noting there and in any 1.1 ↔ 1.2 version-conversion guidance.
• Renderers still implementing the 1.1 external-only behaviour will differ from 1.2 here; the ASWF/OpenUSD normative docs should record which semantics they target.

Further corroboration: the OpenPBR companion report

The companion report — Portsmouth, Kutz & Hill, "OpenPBR: Novel Features and Implementation Details" (SIGGRAPH 2025 Course; arXiv:2512.23696) — states the same 1.2 semantics and, usefully, the rationale:

The specular_color parameter tints the Fresnel factor of f_dielectric, but only in the macroscopic BRDF, i.e., the light incident from the upper or lower hemisphere which is reflected back into the same hemisphere. The light transmitted from above or below is assumed to be unaffected. … This was deliberately chosen to differ from the behavior of the specular_color parameter in Autodesk Standard Surface, which had the tinting but the light transmitted through the dielectric was then compensated accordingly to preserve the energy balance, thus generating a complementary color if specular_color is not white. This was generally considered a bad behavior, hence we opted to fix it in OpenPBR.

So the whole-BRDF tint — energy simply removed, with no transmission compensation and hence no complementary colour — is the intended, documented 1.2 design. This reinforces the target for the MaterialX fix: tint all reflection (including TIR) and leave the transmission untouched (no compensation).

[sherholz]

Verified in both:

• the OpenPBR reference graph (reference/open_pbr_surface.mtlx), and
• the shipped MaterialX library (libraries/bxdf/open_pbr_surface.mtlx, MaterialX 1.39.5).

@portsmouth I am wondering how you validated the behavior of the specular_color tinting also the internal reflections?
Did you mostly looked at the graph or actually tried it out in a renderer?

I am asking because this is also the behavior I would expect when looking at the graph and when I tested it in Cycles, when using the Honey example but with a red specular_color. Interestingly Karma, which also claimed to have implemented everything according to the MaterialX graph behaves differently:

Spec	MaterialX (Cycles)	Karma

Having another example matching the expected behavior of the graph as validation would be helpful.
Could you an Arnold rendering of the Honey example with red specular_color to compared it against the others?