These models give quite good statistical results:

https://huggingface.co/Anzhc/Anzhcs_YOLOs/tree/main

They work both in bbox and in segm.

For XL detailing, the Anzhc Face seg 640 v2 y8n (Anzhc Face seg 640 v3 y11n) models works statistically well. It detects faces quite confidently even from partially side angles and at tilts.

The result of detecting medium and large objects can often be improved by connecting the sam_vit_h_4b8939 model to the detailer node in the CUI.

@Cyberdelia Thanks for answer, but the question is why, the reason :)

@H4rryM3ss influences how the clip model weighs the prompt vs being creative. If you place it between the loader and your prompt, it will diminish the strictness of the checkpoint, letting your prompt come through more. Same with lora, the lora will attain a more weighted output in the conditioning

@Cyberdelia HA!! Now we're talking!! Thanks for the quick reply. I LOVE your work!!

@Cyberdelia It would be cool if there was a prompt tool that you could tailor toward Pony or Noob or SD3 or Flux... I guess I could build that into the request... Hmm... let me play around with that a bit... 

What is this helper tool and how do I find it?

@abrah27861 There's a link at the bottom of the model descriptions. But here it is:
https://chatgpt.com/g/g-6834133e3ab881918a91b3ec6b9eb01f-cyberrealistic-prompt-helper

Since 8.5 there is little step forward to realism, but every step costs more and more adherence. I think, there isnt much anymore in Pony6 tuning.

@berengersimon871 Yeah, I've been playing with 12.5 since making this post, and I am noticing the increase in detail here and there... granted, I'm still not entirely sure the juice is worth the squeeze over older versions, but I am enjoying throwing old prompts at it to see the differences (until I get frustrated and switch to a more forgiving version/another checkpoint altogether, anyway)

@louisisgo Yes, but this is still the best balanced realistic model. I always come back.

@berengersimon871 
I like this series of models, and have been using it almost exclusively for the better part of a year now, while at the same time trying others now, and then. Lately though, I've found that Babes by Stable Yogi v5 is hitting a sweet spot for me, and I've been doing everything with it.

Why?

@Cyberdelia Why important? Whats the difference in prompting?

@greytonbryer280 This how the Helper described the changes:

The model has shifted to a more semantic, context-aware style, less focused on raw tag spam.

Score tags still work, but they should be cleaner and more intentional.

Emotions and actions often work better as sentence-like phrases, and over-weighting (e.g., :1.9) can sometimes make things muddy or cause "mushed" anatomy.

The new model is also more sensitive to pose and detail conflicts, so we have to smooth that out.

@JDHPics1212 Im aware, more xl prompt work, and less booru. But how chatgpt does it know?

@greytonbryer280 It's a custom GPT model.

"Considering as how this is largely a model for use in porn image generation" is a huge assumption. I mean, true, but still.

If you look back at older comments, you'll see v12 was widely viewed as having a colour tone that was seen as much too cold, and intense. This resulted in images that were too bright, and leaned more into white, and blue tones, while people wanted warmer colours. v12.5 dialed that back some, but it is still brighter, and colder than say v10.

@v12h CyberRealistic Pony v12.5 has a corrupted or improperly merged Encoder1 (text_encoder/model1).

Encoder2 is intact, which is why your generations still work — because SDXL usually defaults to Encoder2 at inference.

This occurs when someone merges a model with an incorrect FP16/FP32 format or tries to manually inject a CLIP encoder

@klaabu

> because SDXL usually defaults to Encoder2 at inference.

wow! i never knew this. i'm trying to enable longer prompts than 77 tokens. i've been recommended by AI to concat the prompt_embeddings of encoder1 and encoder2. But i'll tinker around with using values strictly from encoder2. This was extremely helpful and let's me have new alternatives to what I've been bashing my head against for the last few days.

Thank you so much.

@v12h feel free to reach out with any questions :) I also have discord

Check out @Felldude work and articles, you'll learn a ton

@klaabu i've joined!! thank you so much.

Yeah it's possible. What you have to do is find a black guy and get him to put his beef in a black girls taco.

@Breezy00  lol

Try: african woman, ebony, dark skin
And if they are still too light, just add these to negative: pale skin, light skin, caucasian

@EmpireOfSteel_Stahli  worked, thank you

What's your method for training your character loras for this model? I'm not having great results.

I made a detailed video that covers every aspect of my training technique. It's long (90 minutes), but when you have the time, it'll provide full context, and fill in any gaps in your knowledge that you might have. This is assuming you're training locally on your PC: https://youtu.be/UFWjh4iyvEo

Generation3dX oh wow. I was hoping for a little tip and got the entire tutorial. Thank you!

GasparGames honestly, it's the only way to really answer the question, as there are just so many aspects to doing this kind of work that you need to get right. Hope it's useful!

Cyberdelia yep. This model trains beautifully!

Cyberdelia oh this one? https://civitai.com/models/312530/cyberrealistic-xl - tried it and my girls trained on CRP don't come out looking like them. But perhaps this is to be expected. I can try training one of my LoRAs on XL and see...

Generation3dX Using Daz to generate the training images is a great idea too. Thanks for the help man.

Can you elaborate? I'm not seeing any particular issues with 12.5. What resolution are you using exactly?

latina and asian simple work. Black girls are "dark skinned female"

dark skinned will do that regardless of race. so no. africa, african etc work. remember, certain features (hair, freckles etc) will override the race preventing you from making what you want.

davontaxavy109 Although both terms are a bit dated, the word you want is "Negroid." This was the scientific term used by anthropologists in the past. "Negro" was always slang, in English at least, and has since become offensive due to the history of who was using this slang term.

It should also be noted that where you position prompts can have an impact. For instance, if the first thing you put in when defining a girl was "Latina" it seems to do a better job of generating a Latina girl, than if you placed it further into your description.

Also I find plain language descriptions don't work too well, it's better to describe things in short sort of machine like ways. Then when defining a specific thing, like say their hair, it's best to do it all in one prompt, rather than split it across multiple prompts. Finally, you can really only give your main character extensive descriptions, it gets pretty confused if you try to define anyone else you may have if you tried to go beyond a simple one prompt description for them, and even then could still end up swapping features around.

One other thing is that as it if you wanted it to generate an image of a darker skin girl, having sex with a lighter skin guy, often enough it will give you the reverse. I'd guess this is as it was trained with more images of a white girl, having sex with black guys, than the other way around.

And how does the model behave at high resolutions, such as 1024x1536 or 1280x2370 or more, with 70+ steps?

Not necessarily.

One thing is that you ideally want to use as few prompts as possible, the more you use, the more likely it is it could get confused, and not give you the results you want. Meaning just adding a bunch of extra prompting, because you saw others use them, may only make things worse. Then you should only prompt, the things you want in your image. If you had put in a prompt for detailed hands, but had a solo character with her hands behind her head, because you had specifically told it there should be detailed hands in the image, it may just fail to put her hands behind her head, or add the hands of someone else coming in from the sides. Also sometimes you may want to use detail related prompts, to try to reduce negative aspects of the model, and/or LoRA you're using.

Telling it you want their eyes to be highly detailed, especially if stated in the descriptive prompt for the eyes of your character does usually give a clear improvement in the results, at least if it was a close up of their face. Although if the face was not close up, then it may not make much, if any noticeable difference. Meaning you need to know when to use things, to get the results you want.

Facial structure in Pony is not very variative and you hardly can achieve anything without prompting characters Pony can recognise (some Overwatch or DoA characters f.e.)

But you can learn a thing or two from prompts of those images:

https://civitai.com/images/69559690

https://civitai.com/images/66575235

They are a variation from prompts of some person who was a master of such things, but was unfortunately banned some time ago

try "twins, clone, sister" in negatives, sometimes it helps

you can only do that with regional prompting. Search for that term on how it works.

inpainting the woman's face in post would be a good work around

group their details, make use of cross-attention linking to provide a strong element like 'redhead profile' then follow it with that subject's specifics.