consistent collars, seems to work ok
Description
FAQ
Comments (10)
Master is back!!
How do you feel about the training and results of training on concepts like this on Z-Image compared to the other models you've done similar loras for?
to early to say. and i didn't have time to do A/B testing yet. the wierdest thing i find is that the model produces almost replicas of the training images in early epochs when prompted with its captions. so overfit seems to be a big problem.
@Joschek Yes, I've heard this. Honestly, the fact that we're able to train due to a hack is great, but that result is to be expected until we have an undistilled model. I still want to train for Z-Image Turbo though. It's such a fun model.
Excellent!
Which tools do you use to train LoRas for ZIT?
i train online (mostly here on civit), for dataset prep i have an article, it's a bit old now, maybe ill update it some time
What's the weight you use? Because at 1.0 it produces an entirely flat skin and dull (blurry) image features
as low as necessary :) dont remember what i used for these images but i wasnt 1 probably. z image is quite sensitive to lora strength (as opposed to flux before).. this was my first zimage lora and learning rate wasnt right. , but imin the process of training more :)
@Joschek From playing with zit it seems it's really a bitch when it comes to LoRA. Two LoRA at the same time with too high wight completely destroys zit ability. Recently tried this one here together with Jib Mix Realistic (both at 0.7) and it entirely stopped being able to do text.
@supertyp yup thats how it is. Just go lower. 0.3 or 0.4 still can have the desired effect depending on lora and what you are trying to achieve