[epicycles33]

Interesting article. I think the real question however is whether imposing complex priors (say driven by a neural net) makes images better _on average_ even if has some failure modes. My guess would be that a fairly weak prior trained on a diverse enough dataset would lead to better average image quality(as judged by everyday people in diverse scenarios) and that's why they are used.

[wyager]

I think it could absolutely make images better on average, assuming their prior is at all representative. The question, though, is whether the expected benefit to the photographer in cases where it improves the image (usually somewhat small) outweighs the costs when it screws up (perhaps relatively large).

Now, are most people going to notice that the iPhone wrecked the text on their subject? Probably not. But they probably also wouldn't notice if the model wasn't applied to the image at all. The median consumer probably mostly benefits from (in terms of how much they like the photo) AE, a bit of curve reshaping (using a smoothed histogram CDF algorithm or something), and maybe some extra saturation.