I noticed there is an increasing number of people who really seem to hate txt2img models like Stable Diffusion for some reason, allegedly because it's "stealing" from "real artists" or "taking their jobs".
A large portion of those people don't even understand how such models work and spread bullshit "theories" claiming it doesn't actually generate any of the images, but rather mashes existing images together, that it scrapes the Internet in real time, and the dumbest one of them all, that it is somehow able to fit the entire multi-terabyte dataset it was trained on in a single ~4GB file.
First, let me clarify a common misconception that these people tend to believe - copyright infringement (which is what the people likely mean when they talk about "theft" or "stealing") is not theft and anyone who calls it that is either misinformed or a corporate shill.
With that out of the way, the question we should ask ourselves is not whether such models "steal" images, but whether they infringe on others' copyright, as there is a pretty big difference between stealing an all-rights-reserved painting from a museum and selling copies of it.
Oftentimes, the Luddites will make bogus claims about the scraping of publicly-accessible images from the Web and using them to train such a model being illegal, which is simply not the case. Web scraping is perfectly legal in almost every country in the world and it's done for a variety of reasons ranging from indexing and archiving to gathering data for machine training.
But of course the people who this is directed at won't see it, and even if they do, it probably won't change their mind, oh well...
