UPDATE: I won't insist on this being on the generator. But I'm working on an Illus version!

Alright, this started as a quest trying to get something similar to Sardonyx for 1.5 on Pony because I really like that model, but it always bothered me those baked-in watermarks and the bad hands. So, I trained a bunch of styles with "cleaned up" datasets and I looked for a decent base for hands and style (Bluvoll's 3pxDiffusion). What I got is a decent starting point, I'll keep working on it as I get buzz to keep improving the trainings. I'll share some of the styles in the future too.

Begging part

I have a kofi☕ if you want to support my work, you can tip me some Buzz⚡ to help with the trainings or buy my badge from the cosmetic shop (Four Stars Radar). I also take commissions (characters, people, clothes or backgrounds), so if you are interested on a training, check that out.

Prompting

As I'm a noob doing this, I left a trigger by accident is not 100% needed but it does get the subject closer to what I wanted so is part of the prompt, sorry about the self mention. That will be fixed on next version.

Update: Doing the gallery for a LoRA I discovered that the trigger is a blessing in disguise. Without it, things like anime screencap and retro artstyle work as intended. So, I won't get rid of it.

Positive:

score_9, score_8_up, score_7_up BREAK ict style, * your prompt *

Negative:

3d, monochrome, watermark, * your schizo negative *

The 3d is optional, it changes the style a little.

Sampler/Scheduler:

• Euler a: It gives better hands but flatter colors.

• DPM++ 2M Karras: The other way around, worst hands but more detail. I prefer this one.

I haven't tested others mainly because I'm pretty happy with 2M Karras. Let me know if you discover something interesting with the ones I haven't tried.

Steps/CFG:

• Between 20 and 30 steps seems ok, I prefer to stay at 25, a bit more if details look strange. Or brute force it with hires if you can afford it.

• CFG between 5-7.

Other notes:

Adetailer, face fix or inpainting is a must, unless the face uses like half the image. I'm sorry, heads and eyes are small, and I might have cooked the styles too hard so that's not the best. I'll try to work on it for the next iteration.

LoRAs work decently, especially if they don't have a style baked in (like most of mine 😅), use a bit less weight if that's the case. Also "lips", "nose" and "realistic" can help.

I also tested some realistic LoRAs (look at Emma W image), and they work pretty good without "realistic" in the positive IF they are trained on base Pony. The Pony realism ones will break the style.

Acknowledgements

• @bluvoll: For making the model I used as base and being very friendly to talk about it. Use his models, generate, tip, etc. He spent quite a bit of money on his finetunings.

• @Anzhc: For the crash course on training YOLO models, which ended up helping in cleaning datasets.

• @Yoshi__: He doesn't like the model but was willing to test if it was working ok, so thanks.

License

Whatever Astra says about Pony derivatives and Bluvoll says about 3px on top of that. Please do not upload to other platforms, talk with me first. Comments, chat or discord(indolentcat).

That's it! Generate, leave reviews, funny pictures, sexy pictures, etc. They really help to stay motivated, and they support my work and future trainings! Up to X images can get a pin!