Gather a bunch of SNES images. Downscale them to the NES resolution. Train a de-convolutional (transposed convolutional) network to translate the NES scale to the SNES scale. Then apply this network to NES games to get a SNES-style upsample. If necessary, embed this in a GAN until the upsampled NES games are indistinguishable from the SNES games.

Actually, the resolutions are either identical or very similar, so this is a matter of style. It could work to do a GAN that translates NES style to SNES style, the discriminator guessing whether it is NES or SNES.