I am thinking about creating a GAN learner that is not judged by its outputs directly, but judged by what the output does to some color correction filters to an image. For instance the model would only produce 9 floating point numbers that represent gamma, gain, and offset values to apply to an image. The input would be a 512x512 image and the output would also be a 512x512 image, and the model would be judged on how the modified input matches another image. This could be used to create images that look like they were shot at night for instance.
Folder A is filled with random images
Folder B are images shot at night.
The loss function is something that compares the overall colors from A to B. The model outputs these 9 numbers that modifies image A and then compares to image B with the loss function and then updates the model.
Are there other examples of this type of system?