#image diffusion

Left: Raw output

Right: Completed piece for @arionwind's engagement to @vergess.

I like how this one came out, I'll probably submit it to AWAY. I submitted it!

As always, these images were painted by me, then re-rendered using AI models trained on public domain images and my own hardware at home. No theft was used in training, and no environmental damage was caused by rendering.

#73 AI is Killing the Heraldry Industry or Achievement Unlocked: A Proposal for the Autistic Community concerning Coats of Arms

I have a somewhat outdated but still capable PC with a decent GPU.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!

@arionwind's OC Opallia, daughter of the goddess of the void between stars.

Raw output on the left, 8 hours of inpainting and 6 hours of photoshop on the right.

As always, these images were painted by me first, then re-rendered using AI models trained on public domain images and my own hardware at home. No theft was used in training, and no environmental damage was caused by rendering.

Made with this prompt:

More girl frame characters, this time an NPC cadre of "rival girls," the Apiary Squad. Not pictured is their really terrible handler.

As always, these images were painted by me first, then re-rendered using AI models trained on public domain images and my own hardware at home. No theft was used in training, and no environmental damage was caused by rendering.

Update to this fursona for one of my partners. Used a control net based on the included sketch.

Took about 2 hours of in-painting and correction; honestly really quick. I'm going to be exploring control nets a lot in the future I think.

Anyway, this was the base prompt:

long wavy white beard, chest floof, fluff, sheep, sheep furry, anthro sheep, hooves, hoof hands, white fur, white wool, no markings, fluffy wavy fur, no breasts, flat chest, nipples,

@transfaabulous's Girl Frame character, Fulana de Tal, seen here "harvesting wheat" in an exceedingly racially insensitive propaganda poster made by the evil global corporate government from the setting.

Hints for the white folx:

wheat is not from the right part of the world,

you don't wear sleeveless shirts in the field,

you don't wear a bare head or neck in that sunlight,

you don't harvest wheat like that anyway,

Fulana de Tal is not a real name,

and just because she was a migrant farm worker when literally pressganged into service on the other side of the world does not give her boss the right to even begin...

Listen, there's a reason the other posters in the series show a 14 year old girl Like That lmao. It's the whole setting and it fucks severely.

Play Girl Frame.

Commissioned variant set for @oregoncoastfox's fursona, Jewel. 3 out of 6 images. The more explicit ones can be found at my CivitAI below:

As always, these images were painted by me first, then re-rendered using AI models trained on public domain images and my own hardware at home. No theft was used in training, and no environmental damage was caused by rendering.