@SassyAI For example, I requested women in their mid-twenties, but I keep getting women who are more in their late thirties. And when I say breasts, I don't just mean size, which your model actually does quite well, but rather nipples and areolas, which still look like they came from a Barbie construction set. :D

Edit: But, yes, I know that's still a problem at ZIT in general.

Edit2: I think if you merged your model with the "UnstableBastard" model using "Evolutionary Merging," we would have the perfect model. Both models are super strong in different ways.

@SassyAI Do you change something? Because of the new upload...?!

Any chance we got an Q6 or Q4 so we can use it on a RTX2060 :)

Any news when FP8 or Q8 version will be available?

Yep that's where I started, I used Ms Pink I got the same output so it seems to be working fine at base level. But I was just having problems with getting what I wanted, not sure if it was an adherence or incapacity issue. For instance I was trying to get some thickness and curves and I wasn't getting it. maybe cause I was mostly danboru tag prompting? I mostly used to XL/Pony/ILL so.... maybe ZIT/Flux use natural writing? ty I noticed generations on ZIT are pretty taxing to generate, so obviously you'd expect/want their outputs to be superior.

I've looked around at some of the checkpoints here, do ZIT/FLUX have a thickness problem or is it just coincidence? haha

@TheCAL I also had many issues when trying to generate voluptous/curvy women on Z-Image Turbo. After spending a LOT OF TIME tweaking my prompts I got some better results depending on the seed, but nothing close to what I can achieve with Pony/SDXL/Illuistrious.

This is actaully a limitation of the Z-Image base model.

@SassyAI it would be really helpfull if creators add more voluptous/curvy/fat people in their datasets for training, so we wouldn't struggle so much to generate a non standard body

@snaker07 Good to know its not just me or my setup. Yea if a model can't/doesn't do what you want or doesnt have that kind of creative flexibility then you don't really care how realistic its generations are, lighting, textures etc.

@snaker07 most of the time it's the person doing the prompt vs the model. people say it takes plain text but I have used boolean in the beginning for the subjects followed by brief NLP and it worked better. I've used models that just straight up always make cum out of the mouth or drips all over the body for simply putting in a fluid prompt. Sadly, ZIT really does fall short in some places just as SDXL did with pores.

@TheCAL, 2 months ago I didn't even know how to do a git pull, now I can merge checkpoints after a million times failing. My trick was using Venice AI or Grok, taking a few generation prompts in the posts, and then telling it to find the pattern to make a text to image prompt. Hopefully that helps. v1 is never perfect and we are talking about consumer GPUs here so you'll find flaws. A RTX 5090 is around $4.5K ffs and that's not to mention RAM or SSD prices.

Hey it's been a month since your comment and I see a lot of people "struggling" with prompting.

That's also due to some checkpoints working with tags while others don't and very little documentation (at least on civitAI for most parts).

From my experience all models I tested can do curvy women but some require more extensive prompting.

Some are fine with just writing "a curvy woman" or "hour glass figure".

However if that by itself does not provide you enough...uh "meat" go further and describe single parts of the body.

"A curvy woman with an hourglass-figure, she has voluptous large breasts, a small tummy pouch, (very) wide hips, thick upper thighs, chubby legs and a big round butt. Her overall appearance is soft and round with natural folds and bodyfat".