We've arrived, folks.

This model is designed to generate pictures of OMEGATRONIC, a TF2 bot hoster, who uses the avatar of a robot with a monitor for a head. The model was hastily trained in 4 training runs with my server farm over ~1 hour, after seeing the post™ and looking at , to find over 30 existing submissions.

The model was trained using AutismMix-Pony as a base, extending it with a 128-dim lora. Then, another hundred or so image samples were created, with high quality-ones being cherrypicked and used for the final model. The version you see here is actually compressed to fit under 64 dim, with virtually no quality loss.

Trigger words/tips

The model understands OMEGATRONIC far better than HEXATRONIC or DELTATRONIC. This is likely because my dataset only contained a dozen images of them, so it doesn't know much.

As this is a pony-based model, for best results, use these tags at the beginning of your prompt:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up

You can also use "source_cartoon" or "source_anime" to change the style. Use "3d" if you want the style to be closer to SFM, or put it in the negative prompt to make the style flatter.

Here are some of the most popular tags in the dataset (in order of occurrence)

omegatronic: 182
tf2: 136
robot: 130
team_fortress_2: 117
tv_head: 94
: 78
humanoid: 78
object_head: 78
screen_head: 74
metallic_body: 74
robot_humanoid: 74
screen_face: 74
for_a_head: 71
screen: 71
television: 71
electronics: 71
male: 68
not_furry: 65
machine: 65
: 61
_fluids: 52
: 52
bodily_fluids: 52
hi_res: 48
blush: 45

Additionally, you may use various combinations of the tags, such as "weak omegatronic". Pony is smart enough to understand what you want.

I've found that adding tags relating to omegatronic's body composition like "2 antennae" into the positive prompt tends to give better convergence.

Note that the team_fortress_2 and tf2 flags behave differently, and it may be beneficial in some situations to use only one of them. Pony seems to already have a slight understanding of the team_fortress_2 tag, so it may mess with the structure of OMEGATRONIC.

Limitations

Having worked on research and commercial-scale models, I can definitely say there's a lot of improvements to be made. This is really more of an experiment to see how the community reacts. The image dataset is very limited, comprising of only the existing images scraped off off r34, Reddit, Twitter, and other websites, as well as a good chunk of generated images. If somebody can draw copious amounts of HEXATRONIC and DELTATRONIC art, I'll continue training it to make it even better. Additionally, the model is a bit undertrained in regards to specific bodily features, such as OMEGATRONIC's antennae and slender legs. The model will often get these details wrong, unless specifically prompted to behave or edited with inpainting.

H a v e f u n !

(P.S. OMEGATRONIC, if you're reading this, you really gotta step up the quality of the OMEGATRONIC models. Like, really, it's embarrassingly bad, looks like a 3rd grader's Blender project. It's good that you at least don't use SFM anymore, and I know in the latest posts you at least got the body looking slightly better, but it's still . Hit me up, I can set you up with a proper 3D artist.)