Raw LLM Responses

If you're interested in how AI works, its current weaknesses and tells, and so on, it would be better to actually use AI yourself for a while and explore those limits. For example, your first piece was a hand because hands "are still a weakness." For low quality models and low skill users, sure, but it's trivial to get a small model (typically a lora) to make accurate and detailed hands. You can even inpaint these to improve them on an image you otherwise like, or do another round of img2img with it for a more seamless blend without the need for manual editing. Most people genuinely interested in making high quality AI pieces will learn this in a single evening. People have already commented about this, but poisoning isn't super difficult to deal with for most users. How large models are being trained now, I wouldn't know personally, but the general rule is better input = better model. If someone is making a lora (a small model used to train a concept like an artist's style), they manually go through pictures and sort them based on quality and variety. If someone were to steal your style, it would be with a lora. You can find thousands of loras on websites like CivitAI of all kinds of things, including real people, NSFW concepts, "fixes" for the weaknesses of various common models (e.g. hand, eye, skin detailers), and yes, artists and styles. Most loras are trained on datasets of 20-50 images, but you can do it with less. I've seen lots of people experimenting with single image loras - just one data point. That's a bit unrealistic for your average user, but my first lora was made with just 18 images on CivitAI's default settings and it's great. It allows me to create a consistent AI generated realistic character. The technology is actually astonishing and there are AI tools inbaked into a lot of art programs now already, like Krita. There are social media accounts with fake people and images that people cannot tell the difference between (I have seen hundreds of people actually *argue* that something is real online when it's AI and I've pointed out the errors). If you want to "stay ahead" of it, the only way is to experiment with it yourself. You can do that locally and offline, if you worry about that kind of thing. Fooocus (SDXL) is a good entry point - open source too and can run on meh hardware. But it's not getting updates. Flux is the next big thing, but SD 3.5 Large just came out days ago so we'll see which becomes more popular. If you have the time and inclination, learning ComfyUI would be the best, because you can use lots of different models with it and have a lot of control over the workflow. New models can take over in a matter of weeks. P.S. eyes and a preference for butt chin are still a bigger tell than hands on any mid to high quality AI piece :) But I mostly make realistic images, not stylized.