Subpixel sampling for OptiX denoiser

[shocker.0x15]

Hello,

I have confirmed that the temporal denoiser (kernel prediction) is improved with OptiX 8.1.0.
BTW, is there any preferred sub-pixel sampling pattern for the OptiX denoiser?
e.g. Halton sequence

Thanks,

[dhart]

HI @shocker.0x15,

Good question. There are multiple sub-pixel sampling techniques represented in the training set, and one of them is Halton so you should be okay there. Since this is related to your question, I’ll mention that the training images are usually using box filtering for the pixel filter.

–
David.

[shocker.0x15]

Thanks.

OptiX denoiser looks to be designed for offline renderer, so does it assume that the input is decently anti-aliased (i.e. multiple samples per pixel)?
My original question comes from my wondering how it produces a good anti-aliased image when the input sequence is always pixel-center sampled for example.

[dhart]

The OptiX denoiser is trained on a variety of inputs against fully resolved images, so it will always try to anti-alias simply by virtue of it being a neural network, and being taught to assume the input is incomplete. ;) Currently the OptiX denoiser does work best with at least a few samples per pixel. We are investigating how to improve the quality in cases of a single sample per pixel.

–
David.

[shocker.0x15]

Thanks for clarification!