You're effectively training the network for both when training for denoising as your ground truth images are both denoised and anti aliased. With ray tracing bo

You're effectively training the network for both when training for denoising as your ground truth images are both denoised and anti aliased. With ray tracing both problems are brute forced by the same approach of throwing more samples per pixel at the problem.