For those asking about doing AI art or good AI generators that don't bog you down with restrictions, do you have a video card with at least 8 GB of video RAM and 16 GB of system RAM? Try the Stability Matrix project. Open source, runs locally on your PC, although it can take a small chunk of hard drive (mine is almost 30 GB now). But hey, hard drive is cheap nowadays.
If I may link it...
https://github.com/LykosAI/StabilityMatrixAnd if I can't link it, go to github, then add /LykosAI/StabilityMatrix to the URL, scroll down and you will find a link for an installer based on your OS. The app is super simple to install and run. It has a model browser for checkpoints, LoRAs, and more from CivitAI or HuggingFace. You can choose between several different web interfaces to customize your experience. It hand holds you through the entire setup and running process. However, it still leaves you in the dark on how to configure prompts, properly use LoRAs, etc. But it is 100% free, 100% local on your computer (no back end processing on a website somewhere), works offline, which is a wonder how the thing works. AI image generation is a mathematical algorithm, so only formulaic data is contained in the database. It generates images out of noise, and that process amazes me to watch. I always thought it scraped Google or something for the subject matter you enter into the prompt and merges it together. Nope. All formulaic. You say blue eyes, it knows blue eyes equate to a certain algorithm, and with a random field of noise and a seed number it generates the blue eyes you requested. Now that algorithm was originally trained on data that likely came from the internet, and that is the point of checkpoints. People further train the data with something else. Like there is a checkpoint for generating Asian people that was trained on nothing but images of them. And there are checkpoints for porn images, Anime, and several for people who just thought they could train the model to make more realistic images. LoRAs, or Low-Rank Adaptation of Large Language Models, are sort of like a sub routine. They are triggered by putting their name and weight in the prompt, then, in most cases, adding a trigger word for what you need. Think of them like tweaks for what the base language model can't do or understand. I have one for breasts, called the breast size slider. When I want to modify the size of the breasts on my image, I put <lora:breastsizeslideroffset:#.#> in the prompt. I use a number between -2.0 and 2.0 in place of the #.#. Negative numbers make the image's breasts smaller, positive make them bigger. If I don't want to use a LoRA for a particular image you just don't put the <lora:breastsizeslideroffset:1> text in the prompt and it just sits there happily waiting for the next time you call it, so you can have many LoRAs for many different functions installed and they won't all compete with each other unless you purposely call them all in the prompt.
Now a word of warning about LoRAs. There are some that are trained on... interesting data... and can have side effects. Like some LoRAs to modify the female body were apparently trained by using images of a guy wearing women's clothes, and increasing the strength on the LoRA too high can result in his face (complete with beard) replacing the woman's face in your image. It was sort of hilarious a couple of times, but because I had to weaken the strength of the LoRA so his face didn't appear it resulted in it not having the effect I was looking for. Also, LoRAs are trained using a specific version of Stable Diffusion. A lot of them are for SD 1.5, so make sure your versions match or they may not work.
The final thing I will say for the sake of saving images that are almost perfect but may not be perfect. There is a function called "inpainting" when you do img2img. You can use txt2img to generate your pic from text, send your pic to img2img, then use the inpainting function to clean up things like 2 nipples on one boob, bad fingers, messed up facial features, etc. Just be careful because you need to change your prompt. I made an Asian lady standing in the streets of Tokyo and I had her holding up two fingers in the air winking at the camera. Well, like SD always does it put an extra finger on her hand. So I used inpainting to highlight that part of her hand but forgot to change the prompt. So it turned her hand into another Asian lady. Be specific on what you want the inpainting section to fix.
Final thing. Because it is getting hard to tell the difference between AI art and real people it might be beneficial to create a stamp for your art to watermark it as AI art generated by XXX. If your generator creates a PNG file some of them will put the prompt data in the image metadata, but mine does not seem to do that. We don't need AI art floating around Google feeding other AI art systems. Eventually it will be garbage in, garbage out. (Yes, I am bad about this, too, as I don't have a good image manipulation program that can do transparency)
Happy AI art generating!
Punchy