malcolmrey

/

klein9

like 4

ZImageTurbo is easier to train, though Klein9 is not as bad. However I have two datasets that trained well in ZImage but less so in K9 so I'm investigating why.

I would say that there is no clear winner and right now there are pros and cons for both.

ZImageTurbo

• excellent training and likeness
• acceptable lora size for the quality it represents and very good training times
• very good consistency in generation (rarely an output "fails" to deliver)
• speed and quality
• high community adaptation

• still no BASE model so we are bound to "one main model", i disregard the "finetunes" as they are no different than just a model + a lora
• no lora stacking (as a result of distillation), impossible to stack multiple lora of the same character to boost likeness and/or add any other style/clothing/etc loras on top

Klein9

• very good training and likeness for lora
• excellent likeness in training lokr but at the cost of size, still, worth doing if one really needs it
• speed and overall quality
• has base model so it is finetuneable
• better inpaiting (this is purely subjective but it looks like the inpainting is superior here)
  +- the community adoption is not there yet but it is ramping up
  -+ probably user error, but lora stacking on distilled model is not so great (though still better than on the ZIT)

• typical flux randomness, if i hate something about Flux family then it is the output randomness; even if you have a really good lora and you know it because you have generated a lot of cool outputs with it, then you get a random generation that is nothing like the model
• tendency to generate additional limbs

It is good that both are available, Klein is still young and I even did not test the editting abilities of it :)

Thank you very much for your detailed and comprehensive answer. Well then, I'm very excited about the Z-Image Base model and, above all, whether it has the same character consistency as its little brother.

https://www.dropbox.com/t/uadylRRA3kcyilOF images for Sarah Chalke

For me Klein is much superior, the results are absurd, I hope you continue doing news loras.
OBS: If you accept recommendations I would like to see jenna ortega and zendaya

For me Klein is much superior, the results are absurd, I hope you continue doing news loras.
OBS: If you accept recommendations I would like to see jenna ortega and zendaya

Based on the the likeness? And if yes, based on Malcom’s Lora’s? And if not, did you make your own? And if yes, would you share your yaml? Because I haven’t seen any Klein character Lora’s better than ZIT. Please prove me wrong.

+1 to what @Creator6278 said, for me ZImage seems a bit better in likeness too. So if you have different settings, I would love to see them, @printbru ;-)

For me Klein is much superior, the results are absurd, I hope you continue doing news loras.
OBS: If you accept recommendations I would like to see jenna ortega and zendaya

Based on the the likeness? And if yes, based on Malcom’s Lora’s? And if not, did you make your own? And if yes, would you share your yaml? Because I haven’t seen any Klein character Lora’s better than ZIT. Please prove me wrong.

I used Malcolm's LORA with referencelatent node. I don't know if there is this feature in ZIT, but it works very well using real photos of subject. I'm also using donut-nodes for Lora. I think it helps, if you want I can send you an image I made. I've never done anything so close

@printbru the reference with real photos is a good idea and it should definitely help so it is a strength of the klein model :)

dont know what are donut nodes so please link it, and you can also link the image so we can compare :)

@malcolmrey here my friend https://github.com/DonutsDelivery/ComfyUI-DonutNodes.
https://pasteboard.co/Jlw7dTkNugko.png

for me is good. right?

@malcolmrey here my friend https://github.com/DonutsDelivery/ComfyUI-DonutNodes.
https://pasteboard.co/Jlw7dTkNugko.png

Your image is gone but I was able to see it for 5 minutes. As Albert Einstein used to say: Everything is relative. I assume that your image is supposed to depict Scarlett Johansson. Z-Image is far better than this. Have you seen Malcoms example pics?

assuming you want a close-up. Maybe. but from the tests I did I really didn't get as good a result as klein in just a few seconds
and it is also easier to customize the style of clothes and hair.
in my opinion

assuming you want a close-up. Maybe. but from the tests I did I really didn't get as good a result as klein in just a few seconds
and it is also easier to customize the style of clothes and hair.
in my opinion

Why don't you upload your image again so that others can judge this. And I will make a similar image in Z-Image as well tomorrow at the latest to prove you wrong. Your image reminded me of XL characters which doesn't even come close to ZIT.

Why don't you upload your image again so that others can judge this. And I will make a similar image in Z-Image as well tomorrow at the latest to prove you wrong. Your image reminded me of XL characters which doesn't even come close to ZIT.

hey chill man. here https://pasteboard.co/OdQNfH40ZKoO.png
do your best

That's not the picture you've uploaded before

Why don't you upload your image again so that others can judge this. And I will make a similar image in Z-Image as well tomorrow at the latest to prove you wrong. Your image reminded me of XL characters which doesn't even come close to ZIT.

hey chill man. here https://pasteboard.co/OdQNfH40ZKoO.png
do your best

That's not the picture you've uploaded before

yes, the first one I did with just 6 steps and upscale, this next one I used second pass

Was it good for you?
If you have some lora you want to share I would appreciate it