I've been experimenting with the limits of Lightroom's new Generative Remove feature. As with all image manipulation technology based on machine-learning that I have used to date, there are indeed limits—some probably caused by the particular contents of the samples used to train the neural network and others, I suspect, that are an inevitable result of manufacturing data to fill in the "holes" left by removed image elements.
Still, with a little work, I've been able to produce results that are quite plausible from captures that previously seemed essentially hopeless. I've attached before-and-after renderings of a photo of a street crossing in Tokyo that was marred by distracting light and traffic signals as well as pedestrians who were cut off at the edge of the image frame. This strikes me as close to a worst case because the image includes people occluding other people: in some parts of the image you don't have the type of patterned background that would simplify the process of generating the fill-in data.
It took a bit of fiddling to coax the neural network to recognize what exactly I wanted to remove and to paint acceptable alternatives into the image. I also had to segment some of the removals and perform them sequentially, as is sometimes necessary when using Photoshop's Content Aware Fill, but the process was reasonably straightforward and the round-trip to Adobe's servers was commendably swift.
At 1:1, I can identify some artifacts, but for online posting this seems like a decent result. Even an 18x12-inch print looks fine at a reasonable viewing distance.
