www.pugetsystems.com
nvidianews.nvidia.com
wccftech.com
It is physically based.
It's about as much "physically based" as those other reflectance models much like Blinn-Phong where lighting artists are forced to constantly tweak the scene lighting to get their desired reflection output prior to the widespread introduction of the more physically grounded shading models during the last generation ...
It is not since the published paper immediately concedes this point at the very beginning of section 9 ...
Just because the result is ray traced does not necessarily mean it is qualified to be "physically based" much less physically accurate. In fact doing ray tracing exposes the non-reciprocity flaw of the neural reflectance model more easily since the incoming/outgoing direction of the rays means that the lighting results are directionally dependent(!) in a scene. In a physically based shading model, all reflectance models are supposed to respect the Helmholtz reciprocity principle ...
It's as "physically based" as the shader graphs from which the MLP was derived. Loss functions ensure it adheres to the law of energy conservation, and even if they didn't, one could simply apply clamping to the network's output to maintain the energy conservation. This comparison to Blinn-Phong has no basis in reality and and is made up.
What traditional shading algorithm has complete reciprocity taking into account numerical precision? Given the mountain of hacks rendering will always be, complaining about reciprocal approximation errors of a compression/interpolation system seems a bit premature without comparing it to all the other error sources. I doubt it matters.
Shader Graphs have absolutely nothing to do in relation with physically based rendering as they're an artist material shading composition system whereby shaders are constructed with a node-based graph interface instead of an explicit coding interface. Also your idea of 'clamping' the neural network won't really achieve anything since their main problem is that there very few methods to integrate their outputs from them ...
If you're going to marginalize the lack of both energy conservation and reciprocity properties in comparison to numerical precision issues then I have the impression that you don't seem to understand the fundamental theory behind physically based rendering. An inaccurate model will generate incorrect results for some given inputs as opposed to an imprecise model where the least significant set of digits contribute to a much smaller amount of error in the result. The industry has moved on to physically based models for a very good reason since no amount of precision can fix prior problems with the older reflectance models ...
What's the limit on error (and bias thereof) which turns it from imprecise to incorrect?
Artists can still specify inputs similar to that of physically based models but the lighting/material interaction evaluation is inherently invalid on deeper grounds because of it's very theory being wrong. Just because we use all sorts of approximations in real-time rendering doesn't serve as an excuse to straight up regress on our currently more rigorous and correct theories with physically based rendering as opposed to some euphemism like neural rendering ...
They do, because graphs also define the emissive materials, so any neural network training on emissive materials would require the NN to learn the emissive properties of the material as well. This cannot be considered as breaking the conservation of energy.
There is no regression, the purity of PBR isn't what created progress, it was the predictability for the artists. The paying audience don't care if a PBR renderer can analytically go to white in a white furnace test at convergence. As long as it's white enough, it will be predictable enough for the artists and look good enough to the audience.
An emissive material is just a surface that also serves as a light source, nothing more and nothing less. They pose no relation to a neural network's guarantee that their generated output is Integrable or not ...
@Bold Well it doesn't help your argument that the original paper is coming out against your claims again ...
I don't know how many I have to repeat this to you but this generated approximation from the neural network can't be mathematically integrated! Also the normalization workflow they mentioned is unimplemented and would be slower which defeats the main benefit of the original paper which was faster material evaluation ...
The foundations of PBR such as energy conservation and reciprocity did indeed make art more predictably physically grounded. I don't understand why you're passing off a working theory as an insignificant subject when it serves a the core basis of what we're doing today in games ...
