Hi again,
This is my first approach to LoRa Training on Zimage deturbo. Is still in alpha stage and I will improve it when the full Zimage model is publish.
This LoRa is a bit overtrained so keep the weight between 0.8-1.0. 0.9 works reasonably well. I use res_2s sampler and beta scheduler but fell free to experiment with it. Trigger Word: cum on face
example prompt:
((cum on face)). pov photo of a Italian woman age 20 with cum on face , on her knees in the bedroom floor wearing a blouse.casual amateur photo of her face taken from above with natural light. Have fun!
Description
FAQ
Comments (10)
Yay! Your model is so great on SDXL, can't wait to try this on Z-image!
Not sure how is gonna play out, I have new ideas about how to approach the Zimage training but be patient my training rig is garbage and takes time :D
@alejnd77 what are you currently using?
@PixelPerv 4060ti, the bare minimum for training
@alejnd77 How long does the training take with that card? Also I'm in process of setting up Z-Image in comfyui, does this lora work well with character lora's? or does it change the face too much?
@Bliney currently you cannot combine zimage loras without destroying the model, for that we need to wait to train on the full model (not released yet), not the turbo version. Also in my case a training can go as long as 10hours, so any mistake is painfully slow (and expensive) to correct.
@alejnd77 when you say full model, you mean something else than something like this? https://civitai.com/models/2196015/z-image-de-turbo?modelVersionId=2472641 if that's the case, can you give me link to more info or something, so I know what to look for? Thanks!
@Bliney thanks, but I mean the actual model that was used to create the turbo version and is yet to be released
Great work! Any plans for a z-image base version?
yes, but sadly I don't have the hardware to do it right now