OMG a perfect gaussidorov score!!!

@Cyberdelia Хочешь проверять свои проекты напиши сложную подсказку и запусти генерацию 10 изображений и смотри после генерации что получилось , если вам не нравится мои оценки я больше не буду писать у вас в постах

4 месяца назад я ставил вашему проекту 10/10 если что

@Cyberdelia  Тогда я тоже прошу прощения не хотел тебя обидеть если что :)

@gausssidorov928 What that means, its perfectly color calibrated? My monitor sucks than.

@Cyberdelia Yes, i used to it and now like Nosferatu white girls, but i use Vectorscope(forge, auto extension, where you can calibrate the parameters). So, i will give another try. Is there something beside color correction?

@Cyberdelia Now! My monitor really sucks, or just me, who get now blue tint?

@Cyberdelia Allright, with minus blue and plus green and red its much better:) Better colored, than 11. Thanx.

@zyrionteyon your monitor is alright. Yes there is a blue tint because color temp is wrong calibrated, but this kind of mistake is normal and it should be expected. Color correction is hard and finicky. We hope it will be fixed. Still, the colors are much better than 1.0 so it's an overall improvement. I'm sure 1.3 would be amazing when color temp is corrected to a more neutral colors/warmer tones. It's just a slight deviation but it's gonna make a world of difference in things like skin color.

@ElectricDreams Its not just 1.3, but every model now:)

@Cyberdelia Or i just used to the previous models. I even cranked down the saturation for v11, but this is another level:)

i think there's a push/pull going on in general - MOST people want 1-click goodness and a pic that 'pops' right away, so creators must go colorful and contrasty to appease them. SOME people (wisely) want less contrast and more natural colors/lighting so they can be manipulated in editors or be finessed with longer AI workflows, and not have to lose pixel data the moment they make a further adjustment (editor) or start overcooking i2i images (workflow).
i'd be curious if i'm even kinda right according to cyberdelia's thoughts, i noticed they made an Alt1/2 for version 9, seems that might be the attempt at appeasement i'm speaking of, which would balloon (at least some parts of) the process of supplying models by 2x.

@Narz I just want more natural colors on checkpoints. And It's a lot more versatile in the long run: If you don't like natural colors, you can change that very easily, later on. What you can't change is a checkpoint with oversaturated palette, overblown lights and such. That already imposes a personal bias, a personal aesthetics over the user you can't easily change. It's ok if your checkpoint is named "super obscure gothic dark violent XL" but not for a checkpoint that strives to be generalist allrounder. Besides, if you want to train loras on a model, you want one with color neutrality.

The new v12 seems really good so far, but v7.0A and v9.0-Alt are probably still my favorite CD-CR Pony models. I'm hoping someday all the branches will eventually merge back into something with SDXL image quality, Illu adherence and Pony flexibility.

@Cyberdelia 
Seriously?

When will you be releasing new ones daily?

sure but not all will work, depend on what the lora was trained

@carlosurportable419  im trying to create realistic charater Lora (female face)? Any tips when using CIVITAI lora training? Settings etc.

@Bach_e sorry nope, you gonna have to try and try, i only tested local, but no good results for now

That's probably because half of the top 10 are mine, and are all based around similar prompts. Really not that hard to get diversity in the faces if desired IMO

@AIscendant please advise ways to add variety to faces

@turtle03391 Easiest way is to specify an ethnicity/nationality; e.g., French, Italian, Swedish, etc.

Another good way is to describe specific facial features; e.g., wide nose, high cheekbones, freckles, etc.

I also find that different hairstyles, clothes, and even things like styles or lighting will affect the face. Finally, there are no end of LORAs that offer a variety of different face types/styles. Hope that helps!

@AIscendant I'd heard specifying a name in the prompt (like, "marjorie/annabelle", any name not belonging to a famous celebrity or pop culture character) differentiates this. Sounds like an easy solution.

But I just tried and it's inconsistent. Anyone know more about this fix's method?

Agreed (v11 here). Having tested about 8 or 9 others, if you want realism this seems like the best. Unclear on changes for v12...

use your negatives, penis, 1cock etc

Try positive, "Females Only", and negative "male, man, boy, cock, dick, penis"

This model gravitates towards porn/sex, then some LoRA you may include, like ExpressiveH, also make it so you'd be more likely to have a sex scene. Some poses, and view points would also make it more likely that a dick will pop up, as it expects sex.

If you just want just one character, you can also include prompts like "alone," and/or "solo," giving them a high prompt weight can also help. Negative prompting I find hit, and miss, but in general what I've found with generative AI image creation is that our prompts are seen more as suggestions, rather than instructions. Often enough you need to mess around with how you prompt things, and just make a bunch of images, until it decides to comply.

You also need to know that it just can't, or really doesn't want to do some things. For instance, I've tried having a girl kneeling on the ground, giving a guy a blow job, who is sitting in a chair. About 95% of the time, if anyone is in a chair, it's the girl, then even if the guy is in the chair, he's almost always distorted, as he may be missing his legs, or something.

Add the node "CLIP set last layer", set it to -2

@Agimax  ty mate, appreciate that

i mean its inception was an influx of toons, anime, manga.... idk why people started calling any of it 'realism', it will never actually apply.. Pony 'Semireal', tops.
edit: this isn't to do with how obvs someone known for 'realistic' models would incorporate it into the model name lol - separate entirely