I've seen some prototype compression algorithms that work like: 1. Heavily compress image 2. Use neural net upscaler to upscale back to original size/quality

I've seen some prototype compression algorithms that work like: 1. Heavily compress image 2. Use neural net upscaler to upscale back to original size/quality 3. Record the diff between the real uncompressed image, and the upscaled image. This should be small if your neural net is good. 4. Store the compressed image + the diff. This is a lossless image compression algorithm. Part of the trick is that the neural net size is not included in the compressed file size, because that is amortized across many different files you'll be compressing/decompressing. It's not super portable, though, because whoever is opening your file needs identical neural net weights to get the identical output to apply the diff to. It wouldn't surprise me if eventually one of these neural nets gets baked into a standard somewhere, and every browser ships with that network baked in. Call it NeuralPNG or whatever.