I wonder if there is a existing tools that do AI denoising of photos like this. It seems like a trivial thing to produce data for, since all we need is denoised photos, and we can add all kinds of noise to produce pairs.
EthanZyh
CS182 introduced a neural network structure called autoencoder. One type of autoencoder is denoising autoencoder. It takes noisy image as input and encodes it into hidden space. Then decoding the hidden layer gives the denoised version of input image.
sharhar
One important aspect of an auto-encoder is that the internal "hidden space" has fewer dimensions than the input data. This means that the model is forced to perform dimensionality reduction on the data, which helps it generalize patterns and remove noise.
patrickrz
Found this slide to be especially insightful just because you can denoise/add noise in photography editing apps like Lightroom. One can check out https://www.youtube.com/watch?v=qaccekBhS6E for more information
LucasArmand
Can we add a denoising process like this into our raytracer to allow us to render at lower samples per pixel/light source while getting better results?
I wonder if there is a existing tools that do AI denoising of photos like this. It seems like a trivial thing to produce data for, since all we need is denoised photos, and we can add all kinds of noise to produce pairs.
CS182 introduced a neural network structure called autoencoder. One type of autoencoder is denoising autoencoder. It takes noisy image as input and encodes it into hidden space. Then decoding the hidden layer gives the denoised version of input image.
One important aspect of an auto-encoder is that the internal "hidden space" has fewer dimensions than the input data. This means that the model is forced to perform dimensionality reduction on the data, which helps it generalize patterns and remove noise.
Found this slide to be especially insightful just because you can denoise/add noise in photography editing apps like Lightroom. One can check out https://www.youtube.com/watch?v=qaccekBhS6E for more information
Can we add a denoising process like this into our raytracer to allow us to render at lower samples per pixel/light source while getting better results?