#Image Diffusion

[vergess]

Hi...I'm not the anon who asked Reach about getting into AI art, but I'm someone else who's interested in learning. The problem is that everything I've yet found that even remotely resembles a tutorial is like, "How to get rich quick by integrating AI synergy into your developing buzzword and produce art faster than ever at a fraction of the cost!" Which. Is not what I'm trying to do.

I have a somewhat outdated but still capable PC with a decent GPU. I have a basic understanding of python. I have played around mostly with DALL-E a couple years ago. I have a loose understanding of how prompts and negative prompts work. What I do not have is the compsci background knowledge to understand the in-depth writeups about genAI, or a lot of the technical discussions.

I want to run the model(s) locally on my own machine. I do not want to use any kind of subscription or membership service. I don't mind not having a GUI, as long as I've got good docs. I want to learn especially how to consistently make art of the same character (rather than just similar-looking characters). I want to properly learn how to engineer prompts instead of just muddling around by vibes.

Can you recommend me any resources?

Rad! Welcome aboard!

So, there are two "big" programs people like to use for running on their own machines: ComfyUI and A111. However, I personally prefer the way Invoke works.

Ultimately, the UI you use should be whatever is most comfortable for you. However, I will say, Comfy and Invoke are super easy installations. You just download the installer, double click, and follow the on screen instructions. A111 is a little more involved than that, as you can probably tell by the fact that it's on GitHub.

However, most tutorials of decent quality online assume you are using A111, so you may want to start there anyway!

Once you've picked a UI you like, check for that UI's YouTube channel. You'll find actual tutorials on those, rather than the corporate optimization drivel you tend to get from general search engines.

Those tutorials will teach you a lot of the jargon that people throw around, too, which will make it easier for you to use search engines to get to the information you actually want.

Once your UI is installed, it's time to pick a model!

Again, the big place to get your models is HuggingFace, but that's not the one I use. I prefer CivitAI, because I like the feature where you can see galleries of images generated with that model. It gives you a clear sense of what the outputs would look like, and if you click on an image, it will usually show you the exact prompt/settings used to generate it!

I cannot overstate how valuable that is for learning to create your own prompts. Being able to see how other people got their results is THE way to learn, in my opinion.

Now, in terms of creating a consistent, specific character! This is one of the things that people have spent a LOT of time getting genAI to do, because it was very much not the "natural" state of the technology.

There's a couple of ways you can do it, but I find that the most consistent results come from using a LORA trained on the character. Training a LORA can be done on basically any computer that can run genAI in the first place, but it's usually a lot faster to use Big Powerful Hardware from Google CoLab, CivitAI, etc.

Making a character LORA is very easy to do if you're trying to make fanart of an existing character, as you can just train on existing images from the source material. It's a tedious process, because you have to make matched pairs of images and descriptions, but that's sort of the difficult part once you get used to it.

However, what if it's an OC you want? That's trickier, because first you'll need to generate about 75 images (30-150 depending on complexity) that can serve as your training data.

Which means you also need to be able to get those consistent character traits without using the LORA in the first place. So! How????

Textual inversion. That's where you sort of... bundle a bunch of closely related tags into a single file, and then feed the AI that file instead of all the individual tags. The linked guide is focused around realistic looking human characters, but you can use the same techniques for other things (furry, mecha, etc) and for other styles (anime, western ani, cg, etc).

Use your textual inversion to generate your LORA training set, and then get that LORA going!

I hope this was helpful, but if it just raised more questions, please come on back. I love talking about my little computer art toy!!!

[zainosaille]

Raw output on the left, in-painted on the right, no touchups in external software this time.

a white leopard with bright blue spots, greek goddess, very long white multicolored hair, looking at viewer, wings, open wings, furry, anthro, detailed background, detailed hair, depth of field, detailed fluffy fur, volumetric lighting,

[dreamscape-ai]

[cyberimagephantom]

[eroartworks]

How's my tight Negligee? 🥰

[a00zmtisay]

[lifeaintsweet87]

CHECK OUT PATREON FOR FULL SET!

Reblog, like and follow for more thick digital babes!!

[my888-ai-888]

[zainosaille]

Raw output on the left, corrected and touched up on the right.

(big eyes), grey eyes, gray eyes, round glasses, silver glasses, furry, fur, A furry anthro sheep femboy, dynamic pose, high angle lighting, blushing, red face, solo, male, femboy, anthro, twink, sheep, flat chest, narrow hips, white fur, all white, fluff, fluffy, floof, chest fluff, lamb,

[alyssa-ai]

I love going to my suppliers and disrupting them with my outfit 😜

[dreamscape-ai]

I hope you enjoy my Artwork. Comments and suggestions are invited. Feel free to reblog if you wish to. Please keep tags

[cyberimagephantom]

[eroartworks]

How is my maid apron? 😍

[noobmasteria]

Elastigirl

[aiartespresso]

[zayllorai]

Merry Christmas 🎅🤶🦌🎄🎁