Train an AI model to make blurry images sharp, by generating unlimited amounts of training data

[valid remnant]

I just had this thought, because I have an image, which is pretty low resolution and blurry. I tried a bunch of websites to sharpen and upscale it, but the quality didn't really improve.

I am wondering, how these models are trained.
Because it should be easy to generate an unlimited amount of trainingdata.

• Scan the web for sources of sharp images
• Download the image
• Apply slight blurry filter
• Use burry image as input and original as output.

It's a pretty straight forward method, so I would guess, that has already been done or doesn't work for some reason? But I can't imagine, with enough training data, that it's not able to properly make blurry images sharper.
Because I can easily imagine a blurry image as a sharp one in my head.

Thankful for any input

Found ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks.
So it seems like it's a common practice already.
But I'm wondering, if it would be possible to get better results, if you only focus on making faces sharp (thus only using portraits as trainingsdata)

To have a more specialized model

Every model i tried, just makes the image super soft and doesn't really add any sharpness or detail.