Just like Instagram has developed its own aesthetic, so has "shot on an iPhone". I fully agree with you, those images have the typical "surface blur" look that all tiny lens + backlit camera combinations produce.
I'd say it's even worse with video. My phone can make 4K videos, but watching them on a 4K TV is painful. They'd probably look better if you would downscale them to 1080p and then re-upscale them to 4K. They'll be a bit more blurry, yes, but they won't have this artificial flat look everywhere.
Memorizing high-frequency patterns and then reproducing them is one of the very few tasks that AI is actually good at. As such, one can use the raw HDR exposure data from a phone and the noisy raw image data from the sensor and produce naturally-looking high-res images from it. The way that it works is that the AI has memorized the small-scale details from millions of good photos taken with large lenses and proper equipment, and that way it can replace the noisy phone sensor data with its noise-free counterpart.
However, such an AI currently needs 10GB of GPU RAM and it runs for 2 seconds on my 1080 TI for a 4K image. So we'll need another 10x in mobile GPU performance before it becomes feasible to integrate such technology into camera apps. And then of course another 20x for live video processing.
I think these models will drop in requirements pretty dramatically - there is a lot of work in finding sparse subsets of NNs that actually do all the work for example, and embedded hardware is getting a big bump in neural network inference due to massively parallel FMA blocks being added to their DSPs.
Also remember that the majority of academic work is done based on benchmarks that don't include inference time or hardware requirements. As these things start becoming useful without these restrictions industry will start tweaking the cost functions that are being minimised to make it useful in the real world too.
Sadly, I don't have a publicly available link for this.
But no, this is not AI upscaling.
Instead, it is representing the input image in a feature space trained to store natural images and then reconstructs the output from that intermediary representation. If chosen well, the intermediary will store details relevant for high quality images, but discard those specific to low quality images, like noise.
It's called a convolutional autoencoder with bottleneck.
Actually, that might be a pretty nice addition to Adobe's Lightroom cloud service. It already syncs between my desktop and my phone, so why not run some auto-improvements upon import.
But I believe for most users, it needs to be close to instant on the device. Nobody wants to wait 5 seconds to see if their selfie turned out OK.
I'd say it's even worse with video. My phone can make 4K videos, but watching them on a 4K TV is painful. They'd probably look better if you would downscale them to 1080p and then re-upscale them to 4K. They'll be a bit more blurry, yes, but they won't have this artificial flat look everywhere.
Memorizing high-frequency patterns and then reproducing them is one of the very few tasks that AI is actually good at. As such, one can use the raw HDR exposure data from a phone and the noisy raw image data from the sensor and produce naturally-looking high-res images from it. The way that it works is that the AI has memorized the small-scale details from millions of good photos taken with large lenses and proper equipment, and that way it can replace the noisy phone sensor data with its noise-free counterpart.
However, such an AI currently needs 10GB of GPU RAM and it runs for 2 seconds on my 1080 TI for a 4K image. So we'll need another 10x in mobile GPU performance before it becomes feasible to integrate such technology into camera apps. And then of course another 20x for live video processing.