I have been fooling around with the generative fill tool in the new beta version of Photoshop and I am finding it to have great potential. Right now it is working well at removing small items, such as trash cans, hoses, rubbish piles, polls, etc. from the image better than I could and in a matter of seconds too. I have even had success at removing whole cars so long as the car was in front of something with out much unique detail, such as a pavement or wall.
I am also find it is really good at expanding landscape images into panoramas, albeit with limited resolution at the moment. This could help with my backdrop business.
Expanding urban and suburban images though, there is still a long way to go. Likewise, AI really only can work at generating objects, believably, it has ample source material for and if it does not, the failure becomes obvious. For example, in the historic sections of Philadelphia (Dating back to the 1700s), many of the buildings use what is called a flemish brick bond. Although the generative fill can reproduce an American bond easily, perhaps even an English bond, it is really poor at Flemish (and Monk) bonds when removing objects in front of the wall. This makes sense since Flemish bonds went out of style in the 1800s and there are few images available in stock catalogs to train the AI on creating them.
Another, yet surprising, disappointment is generative fill is not great at removing electrical wires in front of buildings, especially if you are looking at the building on an angle. In my experiments it either messes up the siding or the parallax or both.
Of course this is a beta version, and I can only expect it to get better once the actual version is released.
With that said, we do need to get past the copyright issues first with either new laws or at least a few dozen court cases for working precedents to become established. Remember, AI, legally, can not create a new unique piece of intellectual property. Only humans can do that. So, by law, any AI produced work is going to be an infringement against someone's (or many's) copyright.
Although Adobe appears to be taking the high road by only allowing their AI to be trained using images from their stock library, others are not so on the up and up, like Stable Diffusion.
Getty Images sues AI art generator Stable Diffusion in the US for copyright infringementThat lawsuit is for $1.8 trillion btw.
For the amateur, this is probably not a big deal. For myself, and other professionals, who will own the copyright of an image that utilizes AI is a big deal that no one really knows the answer to yet. I realize Adobe going about this in such a way that it will remove the liability of a copyright lawsuit from using an "AI image," but questions around if/how those images can be registered with the copyright office are not going to be answered for a number of years. With many of my clients, this is not that much of an issue. Hollywood though is a cut throat industry, so for my backdrops I need the protection registration offers.
