Anything we can provide is just a drop in the ocean of data they already have. Plus, this was how image generation worked before we got diffusion models (see "generative adversarial networks") and they never reached the level of image quality that diffusion models did.
Huh? GANs, even ones like thispersondoesnotexist are way better than SDXL, they're just hyper-narrow in focus usually, whereas SD can do just about anything