base model or not? because only 12.3 gb?
???
???
base... theres turbo (distilled), regular zimage, and omni-base. kind of like flux 2, flux 2 klein distilled, and flux 2 klein base models.
at least thats what ive taken from the repo pages.
The regular Z-Image is what people usually refer to as base model.
The other base model would be the Omni version.
You can refer their github for the table.
https://github.com/Tongyi-MAI/Z-Image?tab=readme-ov-file#-model-zoo
It is the base model, the Turbo model is not quantized (smaller) than base, it is just had training to get the number of steps down to 8 (from 24-50) and Reinforcement Learning to push preference-aligned visual quality and good-looking outputs.
The Base model looks quite a bit worse than ZIT right now, but will be better to train on (Loras and Full Finetunes)
It is the base model, the Turbo model is not quantized (smaller) than base, it is just had training to get the number of steps down to 8 (from 24-50) and Reinforcement Learning to push preference-aligned visual quality and good-looking outputs.
The Base model looks quite a bit worse than ZIT right now, but will be better to train on (Loras and Full Finetunes)
if it look worse than how they did this?
only tried a few runs, but looks better to me in some ways (more natural), and for sure tons of more variations on each seed.
Probably the biggest benefit is having variations on each run. And it might take a few attempts to tweak the prompt to get what you wanted, and being "natural" not every result looks like a "studio photographer" did the shot ;-)
ZIT is much faster though, and "instant gratification" with each run giving a good results (but lacking the variations), and a bit "studio shot" and "model like poses"
Both have their strengths
if it look worse than how they did this?
I don't know. For me skin looks like this most of the time
Pretty plastic
try experimenting with cfg and stuff. and don't skimp on steps. That is usually the stuff around shiny skin. I'm sure quite quickly we get drowned in guides and tips. (it went whoomp on the Hf model list) .
But variety was zit's biggest downfall and i'm glad that at least is gonna change.
I had quite good success with a few zit trained loras, i wonder if they are a. compatible (edit: NOPE) and b base trained are better for both.