EEVEE-Next: Thickness ouput behavior #120384

New Issue

Clément Foucault · 2024-04-07T21:19:56+02:00

Clément Foucault commented

2024-04-07 21:19:56 +02:00

The EEVEE rewrite introduced the Thickness material output.

This parameter is used for approximating the inner geometry structure of the object without having to rely on raytracing for that. This is currently used for SSS, Translucent BSDF, Refraction BSDF and the nodes containing them.

It is used for computing the exit location and normal of transmission rays within an object.
In other render engine, this is computed either by:

raytracing inside the object and find the correct intersection point (too slow for our purpose, hard to implement currently).
using depth peeling to find the next nearest surface in the view direction (lead to view dependency artifacts from effects that should be view independent like translucency).
use precomputed thickness taking either the average thickness or the thickness in a single view independant direction.

We chose this last option for the following reasons:

it completely avoids the runtime cost of thickness calculation techniques inside the object for each effect.
it allows to model some effect on materials that cannot use other techniques (ShaderToRGB, Blended Materials).
it is view independent and stable.

Model Approximation

For all effects needing it, we consider the local geometry either as a sphere tangent to the shading point and of diameter Thickness, or as an semi-infinite slab of same thickness.

Spherical Assumption	Planar Assumption

Other Standards

Adobe Standard Material

The Adobe Standard Material has a reference to a similar parameter called Volume Thickness.

Pre-computed measure of average thickness of the volume beneath the surface as a fraction of the bounding box width. This can be used by rasterizers to emulate the behavior of material properties that would normally require ray tracing through the volume, like scattering or absorption.

The usage is different (as it is for volumetric approximation) and there is no mention about how the average is computed (uniform cone? cosine lobe?).

However, the parametrization is quite good (object local) as it can be baked into texture maps in LDR and reused across assets.

In addition, there is a Volume Thickness Scale parameter to scale the Volume Thickness after baking.

Output Node Socket Behavior

The output thickness value is treated as local space thickness. This mimics the displacement output and allow for a similar workflow (an intermediate node is used to scale back to world space if needed).

A value of 0 assumes the object is single sided and will only model one transmission event.

Unconnected Socket

An unconnected socket will infer the size of the whole object based solely on its bounding box largest axis length and assume uniform thickness. While this is a very crude approximation, this will match perfectly for sphere objects and give relatively good result for simple objects.

Note that considering the smallest axis might be better as it would allow for a single sided plane to have correct refraction.

This behavior is still opened to discussion. The other possibility is to assume 0 thickness and use one transmission event by default.

Versioning

Compatibility with older file should be as simple as adding a value node to the output and set it to the value that Refraction Depth.

Input Data Baking

Generating good input data is not so trivial. We are reducing the model complexity quite drastically and there is many degrees of freedom and heuristics we can use. In any case, we need to shoot rays in a uniform hemisphere (or cosine, or cone) distribution inside the object.

The thickness can be computed as:

the minimum distance
the maximum distance
the average distance
the cosine weighted average distance.

The distance can be:

the raw ray length (easy to understand, export compatibility?)
the distance to the normal plane (better for box like models)
the tangent sphere diameter (better for organic models, matches EEVEE interpretation)

These options can be implemented as part of an operator or geometry node group.

Thickness Scaling Node

To follow the displacement workflow, we will add a new node that will have World Space and Object Space options, along with two input sockets Thickness and Scaling. It will output the thickness as a Float (as expected by the material output node).

The World Space option will scale the Thickness input by the bounding box largest axis length so that it stays constant with object scaling. The Scaling option will just scale the result of the space conversion.

This allow full flexibility and clarity to amend the baked data and provide compatibility with the ASM standard.

Open Questions

Thickness from shadow

This was the default mode in legacy EEVEE. It was only used by BSSRDFs (subsurface scattering). Now that we support more transmission BSDFs, this interact with all of them. This create some funny looking refraction which are distorted by the shadows (which are not even aligned with the shading).

For now we take the min between the user thickness and the shadow one which creates some artifact because of shadow map aliasing. But if we have a cheap way to pre-compute something more correct than this, then it become rather pointless to have this option. Which would speed up rendering and simplify the usage.

The EEVEE rewrite introduced the `Thickness` material output. This parameter is used for approximating the inner geometry structure of the object without having to rely on raytracing for that. This is currently used for SSS, Translucent BSDF, Refraction BSDF and the nodes containing them. It is used for computing the exit location and normal of transmission rays within an object. In other render engine, this is computed either by: - raytracing inside the object and find the correct intersection point (too slow for our purpose, hard to implement currently). - using depth peeling to find the next nearest surface in the view direction (lead to view dependency artifacts from effects that should be view independent like translucency). - use precomputed thickness taking either the average thickness or the thickness in a single view independant direction. We chose this last option for the following reasons: - it completely avoids the runtime cost of thickness calculation techniques inside the object for each effect. - it allows to model some effect on materials that cannot use other techniques (ShaderToRGB, Blended Materials). - it is view independent and stable. ### Model Approximation For all effects needing it, we consider the local geometry either as a sphere tangent to the shading point and of diameter `Thickness`, or as an semi-infinite slab of same thickness. | Spherical Assumption | Planar Assumption | | -------- | -------- | | ![](https://projects.blender.org/attachments/93ae749f-6d2a-413f-91cb-e4401aa36d23) | ![](https://projects.blender.org/attachments/9cc8c3fb-f5aa-4f55-8fe8-d2fcb7112c70) | ### Other Standards #### Adobe Standard Material The [Adobe Standard Material](https://helpx.adobe.com/content/dam/help/en/substance-3d/documentation/s3d/files/225969597/225969613/1/1647027222890/adobe-standard-material-specification.pdf) has a reference to a similar parameter called `Volume Thickness`. > Pre-computed measure of average thickness of the volume beneath the surface as a fraction of the bounding box width. This can be used by rasterizers to emulate the behavior of material properties that would normally require ray tracing through the volume, like scattering or absorption. The usage is different (as it is for volumetric approximation) and there is no mention about how the average is computed (uniform cone? cosine lobe?). However, the parametrization is quite good (object local) as it can be baked into texture maps in LDR and reused across assets. In addition, there is a `Volume Thickness Scale` parameter to scale the `Volume Thickness` after baking. ### Output Node Socket Behavior The output thickness value is treated as local space thickness. This mimics the displacement output and allow for a similar workflow (an intermediate node is used to scale back to world space if needed). A value of 0 assumes the object is single sided and will only model one transmission event. #### Unconnected Socket An unconnected socket will infer the size of the whole object based solely on its bounding box largest axis length and assume uniform thickness. While this is a very crude approximation, this will match perfectly for sphere objects and give relatively good result for simple objects. Note that considering the smallest axis might be better as it would allow for a single sided plane to have correct refraction. This behavior is still opened to discussion. The other possibility is to assume 0 thickness and use one transmission event by default. #### Versioning Compatibility with older file should be as simple as adding a value node to the output and set it to the value that `Refraction Depth`. ### Input Data Baking Generating good input data is not so trivial. We are reducing the model complexity quite drastically and there is many degrees of freedom and heuristics we can use. In any case, we need to shoot rays in a uniform hemisphere (or cosine, or cone) distribution inside the object. The thickness can be computed as: - the minimum distance - the maximum distance - the average distance - the cosine weighted average distance. The distance can be: - the raw ray length (easy to understand, export compatibility?) - the distance to the normal plane (better for box like models) - the tangent sphere diameter (better for organic models, matches EEVEE interpretation) These options can be implemented as part of an operator or geometry node group. ### Thickness Scaling Node To follow the displacement workflow, we will add a new node that will have `World Space` and `Object Space` options, along with two input sockets `Thickness` and `Scaling`. It will output the thickness as a `Float` (as expected by the material output node). The `World Space` option will scale the `Thickness` input by the bounding box largest axis length so that it stays constant with object scaling. The `Scaling` option will just scale the result of the space conversion. This allow full flexibility and clarity to amend the baked data and provide compatibility with the ASM standard. ### Open Questions #### Thickness from shadow This was the default mode in legacy EEVEE. It was only used by BSSRDFs (subsurface scattering). Now that we support more transmission BSDFs, this interact with all of them. This create some funny looking refraction which are distorted by the shadows (which are not even aligned with the shading). For now we take the min between the user thickness and the shadow one which creates some artifact because of shadow map aliasing. But if we have a cheap way to pre-compute something more correct than this, then it become rather pointless to have this option. Which would speed up rendering and simplify the usage.

thickness_plane.drawio(1).png

25 KiB

thickness_behavior.drawio.png

36 KiB

Clément Foucault added this to the 4.2 LTS milestone 2024-04-07 21:19:56 +02:00

Clément Foucault added the

BSDF	Cycles	EEVEE-Next	EEVEE (legacy)
Translucent
Refraction
Sub-Surface

BSDF	Cycles	EEVEE-Next	EEVEE (legacy)
Translucent
Refraction
Sub-Surface

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

EEVEE-Next: Thickness ouput behavior #120384