Topic: Really Basic Questions

Posted under General

I'm having trouble getting some really basic questions about AI art answered. I'm hoping someone here can help me out.

When I hear about AI art, I immediately jump to some sort of website where you enter prompts or a discord that you have only a certain number of tokens to use a model that may or may not be tailored to what you want.

Do you guys have the models downloaded directly on to your computers so you don't have to go through a third party site? If so, how is that accomplished?

I've read something about A111 being the go-to model for training. How do you do this? Is there some sort of application you need to download?

I'm literally just trying to train an AI on a very specific image set and mess around with it. I'm trying not to break the bank doing it as it's just for my personal use. The only AI I have somewhat experience with is NovelAI but that's more for generating images to go along with stories you're writing. The prompt system is good and has a basic model for furry characters but it's not very intuitive and will only allow you to use one reference image at a time.

I'm want to train an AI on a specific style.

skisnep701 said:
I'm having trouble getting some really basic questions about AI art answered. I'm hoping someone here can help me out.

When I hear about AI art, I immediately jump to some sort of website where you enter prompts or a discord that you have only a certain number of tokens to use a model that may or may not be tailored to what you want.

Do you guys have the models downloaded directly on to your computers so you don't have to go through a third party site? If so, how is that accomplished?

I've read something about A111 being the go-to model for training. How do you do this? Is there some sort of application you need to download?

I'm literally just trying to train an AI on a very specific image set and mess around with it. I'm trying not to break the bank doing it as it's just for my personal use. The only AI I have somewhat experience with is NovelAI but that's more for generating images to go along with stories you're writing. The prompt system is good and has a basic model for furry characters but it's not very intuitive and will only allow you to use one reference image at a time.

I'm want to train an AI on a specific style.

If you want to run AI-things locally, you need a powerful computer. How powerful depends on how long you're willing to wait. You can run everything on a bog-standard CPU so long as you have sufficient RAM (>8B), but generations will take many minutes of time. For faster generation, a graphics card is a must. NVIDIA cards generally have the best software support, but both Intel ARC and AMD cards are supported. Most people run models locally with a "web UI", which provides a convenient interface to tweak generation parameters and greatly eases installation from the internet.

A1111 and SDXL are two popular WebUIs. A1111 is older, and perhaps more popular, but SDXL has more features and better support for non-NVIDIA cards. Just look for their github pages to find them, and follow the instructions there.

Models can be downloaded from Hugging Face and civit.ai (civit ai has all the porn models; a good start is YiffyMix 3.1).

For training: SDXL has built-in support for textual inversion training, which works backwards from a set of images (~20 is a good minimum, but more is much better) to figure out what "words" best describe the character. (They're not actually words, rather a vector embedding of the semantic meaning of the words, but it's basically what a prompt gets turned into immediately). These take up little space, are simple to train, and work well for characters that are basic variations of furry stereotypes (e.g loona or whatever that's basically just a goth wolf)

For an art style, text inversion also works, but less well; typically you want a LoRA. I believe the software for training LoRAs is called Kohya. I have never used Kohya.

  • 1