- May 2, 2018
- 729
- 2,548
I couldn't seem to find a thread for discussing AI art generation, so I figured I'd start one. Tricks, tips, advice, questions, that sort of thing. I only just recently got into it properly and by that I mean local generation on my own PC. So I am just a newbie myself. I suspect we have some real experts on this site and it would be useful to collaborate. I'll throw a few things out there to start things off. [I've come from the future to report that most of what's below is just rambling, sorry.]
AI Models: Do people tend to find a model they like and stick with it or do you dabble around?
I'm currently in the dabbling phase. I find it interesting to try the same prompts in different models and see how the results vary. I probably have too many redundant models at the moment, but I often find myself switching and swapping to see if something else will achieve what I want. One thing that I've found very useful if I'm having trouble generating with a realistic model is to use one of the hentai focus models with heavy booru tagging then working back from there and converting it in a real life focused model.
LoRAs: This is an aspect I wasn't really aware of only having used some free generators before. For the uninitiated, you can think of them like add-ons to the model that focus on one aspect to reinforce more specifically that concept that the model might only have a vague idea about or no idea about. Common uses are: specific people (celeb fakes) or characters, outfits, sex positions, art styles, etc. Going through LoRAs is a real rabbit hole, again I've probably ended up with too many as I just like to try them out. I find them to be a double edged sword in a sense. Sure, they can inject a concept into the model, but they almost always fundamentally change other aspects. They can change faces, nipples, body types.
Inpainting: Hoo boy, does AI hate hands. Too be fair, I've heard a lot of real artists struggle with hands, so I guess it was to be expected. They are complicated things. The savior here is inpainting. This is when you have the generator redo only a small portion of the image. Very good for fixing faces and hands, which often need fixing. The funny thing is that sometimes I'll use the image that actually has the best hands, because I can inpaint the other aspects more easily.
Conversions: My very basic understanding of how the AI works is that it generates a random noise map and as it starts getting even the hint of a picture, it starts turning nothing into something. It's kind of like looking at clouds and imagining they are shapes of things you recognize. But you don't have to start from nothing. You can give the AI a bit of a nudge in the right direction to inform color, composition, etc. The most specific form of this being image-to-image conversion. If you give it an image that is already of the thing you want to see, it has a much easier time of making it. (This is essentially how inpainting works too.) You can limit the amount of creative freedom the model can take so that it only makes a few tweaks overall.
I find this most useful in trying to make very specific scenes in a realistic model by first making it in a hentai model. The anime/hentai sims strongly leverage booru tagging, which is perfect for machine learning. It's true that some of the real models use similar tagging, but I've just found the anime ones get there a lot quicker. There are models that handle both and most models do have some mix, but you can start running into problems of mixed styles. Your real-life people look too fake or your fake people look too real. Then you are left having to negatively prompt against the aspects you don't want to include and it can be a real chore.
Prompts: Man I wish I knew what the hell things were called. I've found that if you want to make good prompts you should brush up on your terminology. We've all been exposed to a great number of styles and techniques, but most of us have no fucking clue how to put that exact thing into words. If you're lucky, you can find some cheat sheets that give you a whole list of terms and examples of what they look like. The other side of this is the AI not knowing what the fuck you mean sometimes or not having the right context. I find this most frustrating with distance to subject. You have to remember that the model is trained on images in it's default size (512x512 or 1024x1024). If 80% of it's training data had a woman at about three to five feet away, trying to have an image of a woman at thirty feet away is going to be an uphill battle. There are some tricks, but I've very often found that directions around distance and angle are only loosely represented most of the time.
Anyway, feel free to share you own experiences, ask questions, or leave some advice.
AI Models: Do people tend to find a model they like and stick with it or do you dabble around?
I'm currently in the dabbling phase. I find it interesting to try the same prompts in different models and see how the results vary. I probably have too many redundant models at the moment, but I often find myself switching and swapping to see if something else will achieve what I want. One thing that I've found very useful if I'm having trouble generating with a realistic model is to use one of the hentai focus models with heavy booru tagging then working back from there and converting it in a real life focused model.
LoRAs: This is an aspect I wasn't really aware of only having used some free generators before. For the uninitiated, you can think of them like add-ons to the model that focus on one aspect to reinforce more specifically that concept that the model might only have a vague idea about or no idea about. Common uses are: specific people (celeb fakes) or characters, outfits, sex positions, art styles, etc. Going through LoRAs is a real rabbit hole, again I've probably ended up with too many as I just like to try them out. I find them to be a double edged sword in a sense. Sure, they can inject a concept into the model, but they almost always fundamentally change other aspects. They can change faces, nipples, body types.
Inpainting: Hoo boy, does AI hate hands. Too be fair, I've heard a lot of real artists struggle with hands, so I guess it was to be expected. They are complicated things. The savior here is inpainting. This is when you have the generator redo only a small portion of the image. Very good for fixing faces and hands, which often need fixing. The funny thing is that sometimes I'll use the image that actually has the best hands, because I can inpaint the other aspects more easily.
Conversions: My very basic understanding of how the AI works is that it generates a random noise map and as it starts getting even the hint of a picture, it starts turning nothing into something. It's kind of like looking at clouds and imagining they are shapes of things you recognize. But you don't have to start from nothing. You can give the AI a bit of a nudge in the right direction to inform color, composition, etc. The most specific form of this being image-to-image conversion. If you give it an image that is already of the thing you want to see, it has a much easier time of making it. (This is essentially how inpainting works too.) You can limit the amount of creative freedom the model can take so that it only makes a few tweaks overall.
I find this most useful in trying to make very specific scenes in a realistic model by first making it in a hentai model. The anime/hentai sims strongly leverage booru tagging, which is perfect for machine learning. It's true that some of the real models use similar tagging, but I've just found the anime ones get there a lot quicker. There are models that handle both and most models do have some mix, but you can start running into problems of mixed styles. Your real-life people look too fake or your fake people look too real. Then you are left having to negatively prompt against the aspects you don't want to include and it can be a real chore.
Prompts: Man I wish I knew what the hell things were called. I've found that if you want to make good prompts you should brush up on your terminology. We've all been exposed to a great number of styles and techniques, but most of us have no fucking clue how to put that exact thing into words. If you're lucky, you can find some cheat sheets that give you a whole list of terms and examples of what they look like. The other side of this is the AI not knowing what the fuck you mean sometimes or not having the right context. I find this most frustrating with distance to subject. You have to remember that the model is trained on images in it's default size (512x512 or 1024x1024). If 80% of it's training data had a woman at about three to five feet away, trying to have an image of a woman at thirty feet away is going to be an uphill battle. There are some tricks, but I've very often found that directions around distance and angle are only loosely represented most of the time.
Anyway, feel free to share you own experiences, ask questions, or leave some advice.