(slight emergency saturation increase, only affects ~3 people who downloaded first)

???

I'm sure I left you all puzzled... Is this it? Did UndefinedUsername solve SD1.5?? Is there really no more space for improvement??? Where else can this go????

FEAR NOT!

I'm here to lift your existential crisis! By presenting a new and improved Omnitron model!

Omnitron-Extra!

HOW CAN ONE IMPROVE PERFECTION???

Well, I grabbed a bunch of reference photography, slammed some auto-generated tags on it and ran it all through Kohya_ss finetuner. 4000 images give or take. Only male of course. Avoided the "turntable" stuff, because it's already overused. Once done, ran genital and style datasets again to remind finetuned model of it's greatness.

Resulting model was still a little foggy on it's past, and kept doing it's own thing. So, I made a lora extract from the finetuned model onto base Omnitron. Intending to make a "morph". But to my surprise, instead full transforming base model into finetune - lora produced improved Omnitron output, without going into brainwashed finetune territory! After some tests - it was set - I merged lora into base Omnitron to produce Omnitron-Extra!

EVEN BETTER???

Pros

- A lot less blurry output.

- Bounce back on landscape ratio imagery (base Omnitron had that ruined because I trained it on 512x768 resolution, for no reason; finetune was trained entirely on 512x512).

- Better anatomy.

- Sharper details.

- All the cool stuff from Omnitron is still there! Genitals, styles, abstractions!

- Interactions are a tiny bit better (it's still SD1.5, don't expect much).

Cons

- Not sure if it has any cons over base Omnitron. We'll find out as I test it (some abstractions look different, from preliminary testing, but still quite wild).

- I guess upscaling can go into cons. Because it's tricky to pick right upscaler to retain base image style. Right now I'm juggling 3 upscalers.

WEW

That's all I got! Good luck and have fun!

Omnitron-Alt

Alternative version to base Omnitron. It's ~12% more Megatron, ~7% more RuggedResolveBaraArt, ~7% more PhotonicFusion, and as such ~X% less Triforce.

Hey!.. Psst!..

Psst!.. Hey, kid!.. Want some AI models??..

First one is free...

It's the newest shit on the street - called Omnitron.

It can do ????? and-... Well, mostly just ?????. Hope you're into that.

More penis??

Of course! You guys didn't think I would leave you all hanging, now, did you??

It's my first tryhard finetune! I took all of my narrow arse loras and smooshed them into one thing! How cool is that?

Is it better than Ultron??

It's based on Ultron. But is it better? It depends.

Whole multi-thousand image dataset was ENTIRELY male characters. Not a single new female image in there. Meaning, Omnitron is even more malecentric than Ultron. In terms of female output it's not any better than Ultron, likely even worse (I'll test it later, maybe).

In terms of male output - Omnitron contains all of my genital loras, my most recent multi-style lora and then some sweets on top. If you're looking to generate variety of male eye candy - Omnitron might just be better.

BIG FAT CON!

There is one big drawback - colors are quite bad. I'm trying to find a good adjust that would make it work, but right now model is all over the place.

So, stay tuned for improved versions, if I find a way (might even take a merge).

No, this is still not SDXL

You still have to inpaint some and latent upscale.

Some technicalities

Like I said, first of all model contains male genitals in many shapes and forms. To evoke - use one of the following:

flaccid penis

cut penis

semi penis

erect penis

behind ('ass' should work, but a part of dataset had 'behind' trigger)

Triggers are preferred to be in either first or second position in the prompt. They could work further down the line (hah, literally), but they were always first in tag files.

Next, you can pick one of trained unique'ish styles:

photo art - (mostly AI generated photography of exceedingly handsome male specimens)

render art - (plastic looking 3d render style)

artificial art - (AI generated artistic imagery; quite a nieche look, might not be everyone's cup of tea)

digital art - (airbrushed looking artstyle, somewhere inbetween render and illustration)

traditional art - (style with destinct artistic medium texture; it can be oil, water color and maybe even pencil - they just mix together, and can be evoked separately, to a lesser degree)

illustration art - (various isolated character depictions that didn't have a distinct style to them, might appear similar to colored pencil)

pencil art - (mostly monochrome pencil art and sketches)

niji art - (AI anime style leaning elongated, oh-so-slightly effeminate face; just don't expect shounen, it turned out very bara)

himbo art - (AI anime style with more muscle and round/square face shape, but even though datasets were very different, it ended up a lot like 'niji art')

concept art - (reference sheet style, that's become very popular since diffusion AI improved)

You can try 2 styles in a prompt or even generate with one style and upscale with another - this adds a lot of options for customization. Style tag should also appear first or second in the prompt.

How much each style works will vary greatly per prompt and how much "juice tags" you're already using. But overall, every style had a unique set of ~200 images in training data, meaning they will alter the final output at least somewhat.

noQs and someAs

- Model has a slightly lower CFG threshold than default 7. A lot of prompts do work at 7, but if fuzz and oversaturation appears - try CFG 5 (4 or 6 might work too).

- tag '????' seems to be quite cooked and often breaks style of image. Likely TN-Triforce legacy. Use '?????' instead, if issues arise.

- trained at 512x768 - there's a chance that inpainting genitals will also work better in this resolution.

- Hires.fix is often better at upscaling artistic images than Ultraupscale script, because it retrains style and texture better.

- Style tags can interfere with genital tags. If you're struggling to inpaint genitals in some style tag - just try without it.

- 'Concept art' and 'pencil art' are both a bit tricky to get intended output. Generally, you want to avoid describing background, or use 'white background' tag. Otherwise, they both morph into some entirely different illustration style.

- There was no pornography in the datasets, only ????? and romantic interactions. Multi-character does seem to work slightly better than in Ultron.

- Hands are also a bit better than in Ultron, I feel. Maybe.

- Pins lora is still somewhat better than this model. Finetuning just doesn't really yield the same kind of "magic" output, as lora training. Perhaps it just takes a lot longer.

- Using genital tags will have a slight detrimental effect on overall image colors. Because.

- I did not do 'concepts' this time around, like in Pins lora (mainly because dataset got a lot larger and I got impatient; I do regret it slightly in retrospect). But some training did occur on it's own - wings work quite well for example (varies per style).

- Dataset images weren't all perfectly matched stylistically - I eyeballed them. Expect model output to be swingy as well.

- Overall, there's still a lot of place for improvement. I'll figure out where to go from here, once I make some previews and see what's what.

That's it for now!

Good luck and have fun!