Neotron

---

If you liked this, check out new and improved Sigmatron!

Sigmatron - Sigmatron | Stable Diffusion Checkpoint | Civitai

---

Closing chapter in Neotron trilogy!

I present you all with the final version of Neotron!

Neotron-Welldone!

Unbound was amazing, but I felt it could still be pushed a little further in creativity department. Neotron-Welldone was trained for another 10 epochs of nudity and another 10 epochs of styles datasets - for better or for worse.

As a result, Welldone is heavier on detail and styles and a lot closer to Omnitron.

I personally will be using it over Unbound, but in the end it's a preference matter.

I will post some previews for a bit, but that aside, I don't really have any more training to add to this model. Meaning, it's a goodbye for the time being!

Until next time!

Good luck and have fun!

---

Good news everyone!

There's been a breakthrough in making Netron more creative and less restrictive!

Neotron-Unbound!

Won't go into details on what was done. But the styles are a lot stronger now, and minor concepts - a lot more responsive!

Now, it's easier for me to make use of the model, because I know what caption files looked like. For everyone else it's just a guessing game.

For now - deepbooru tags are the strongest. You can still use word salad prompts, but if they come out meh - add '1boy, male focus, solo', and it should start coming together.

Minor concepts - I'll look into, and if enough trained well, I'll list them in some way. But in general, it's mostly popular character design concepts.

That's all for now!

---

Yo! Long time no see!

A lot has changed at CivitAI, huh? Things are just not what they used to be...

Fortunately, I brought some old school cool for you all today!

Remember we used to have 2GB models? Like really small, but they could generate images as well? I know, I know, it could be shocking for some of you youngsters... But you'll get to witness one with your very own eyes today!

Meet NEOTRON!

You might know it's elder cousins Megatron, Ultron and Omnitron. Neotron is the newest addition to the family.

(did I mention it was male centric?)

By intent Neotron is Omnitron, but more and better. But you all know how intent goes - like a drunk on a bicycle in a pitch-black night!

tadum-tsh

(that's the sound him falling)

So, what happened?

- Styles happened - 12 of them to be exact!

photo style - self-explanatory.

render style - a mix of some modern 3d render looking art.

artificial style - random obvious AI generated art, not the best thing in the world, but individual images were pretty, so I bundeled them into a style.

chargen style - husbandu ready to train dataset available on this website from waaaaay back.

digital style - random softbrush looking art.

traditional style - style with identifiable painterly medium.

illustration style - colored line art and colored pencil art.

classical style - ancient school nude male body study pencil/graphite art

pencil style - pencil art and sketches

himbo style - dull chin bara anime hunks.

niji style - sharp chin shonen (sort of) anime hunks.

concept style - character concept reference sheets.

Not everything happened as intended unfortunately. I did some increased text encoder training and that mixed up the cards quite a lot. Some styles got a bit too subdivided between captions and tags.

Styles like 'classical' and 'pencil' need specific latent to evoke intended output, aka white background with a monochrome shape. To activate it in full, you'll have to prompt for that.

Deepbooru tags 'realistic' and 'traditional media' ended up being a big players as well. They change some styles drastically.

If you're struggling to generate painterly images, you can try lover CFG to 5.

'Niji' and 'himbo' are very distinct this time around.

I also freed up 'art' tag to mess less with existing training in the model (Neotron is an Ultron finetune).

- Male nudity happened!

flaccid penis

cut penis

semi penis

erect penis

covering - pose is guaranteed, but hands are not.

manbutt - deepbooru still tagged 'ass' everywhere, so you can use either tag.

There were a few other concepts that I tried to train, but they either became redundant or did not work out.

'bulge' - wanted to get rid of "tennis ball bulges", but they just kinda got trained out of the model on their own. Don't need to evoke 'bulge', everything just works as is.

'manbutt, tease' - was going for underwear pull imagery, but, again, it learned the concept on it's own. Just prompt 'manbutt' and describe lower body wear - generates pants/underwear around thighs. You can still try to add 'tease', but it has little effect - might sometimes raise clothes slightly higher over butt cheeks.

'root penis' - bas obviously going for semi-concealed male genitals, buuuut model just learned the concept in it's own way again. Prompt 'pubic hair' - penis root is near guaranteed to show up. You can still inpaint using tag 'root penis', if you need to adjust detail.

'naked' - tried to train naked tag to generate 'convenient censorship' imagery, but it just flopped, because neotron expects you to input a pose to generate a pose.

There were also manbutts with 'thong's.

- Upscales amazingly!

- Inpaints hands, genitals, anything!

- Quite consistent!

- Strong prompt adherance!

(entirely unintended, I increased text encoder training to bake-in triggers harder)

- Generate at 1024 px if you want!

(to varied success; model was trained on 512x512, so I recommend SD1.5 resolutions + latent upscale)

And what didn't happen?

- Omnitron's abstract beauty didn't happen.

Neotron still understands abstract very well, but it doesn't go too crazy with creativity.

- 'Male abc' hardly happened.

Trained 26 male facial variety triggers for every letter of the alphabet as 'aaa male', 'bbb male' and so on, very early in finetune process.

For most, only major facial/body features survived to the end.

Can still use them and remix faces if you want, but difference is often very mild and can have detrimental effect on output.

- CFG and Negative Embeddings acting up.

Due to how much prompting style changed for the model, most of my old testing prompts need adjustments to not cause CFG drift.

Influx of fluff tags like 'masterpiece', 'best quality' and whatever else seems to be often detrimental.

And since text encoder shifted, negative embeddings might also have unexpected effect on your output.

Worst case, you can try CFG 5-6 instead of default 7, if you want to bother.

- Prompting got a bit Fluxified.

Xcomposer excessive tagging did a number on text encoder, so much that detailed prompt means better image. It doesn't just dispense random images like Omnitron, it's a little more stiff in that regard.

It can also be very sensitive to random tags, that had no influence before, like 'blurry' somewhere near the end of 2 line prompt, will generate a shaky hands photo.

- Minor concepts barely happened.

'wings', 'horns' and such were tagged to train various fantasy concepts, but once again, they trained in the weirdest way imaginable. You have to prompt them multiple times for them to activate consistently, the more the better it seems - do like 'wings, wings, wings, wings, wings, wings' at the end of the prompt to get wings. No idea what's the optimal number of 'wings' is, don't ask.

Anything else?

I'm not entirely satisfied with the model, as I prefer "pretty image dispenser" models like Omnitron, so I will still try to make Neotron a little more interesting. But it is what it is for now.

I did mark "don't sell", but honestly - I don't care. If you know suckers who will pay you for free crap from the internet - have a go at it.

Good luck and have fun!