Nope, you can't train a good diffusion model from scratch with just a few thousand images, that is just delusion (I am open for examples though). Adobe Firefly is a black box, so we can't verify their claims, obviously they wouldn't admit, if they broke copyright to train their models. We do however have strong evidence, that google, openai and stability AI used tons of images, which they had no licence to use. Also, I still doubt that all of the people, who sold on Adobe Stock either knew, what their photos are gonna be used for or explicitly wanted that or just had to accept it to be able to sell their work.
Great counterargument to my first argument by the way ๐