NOTE : Base model : Z Image Turbo & Flux1
Showcase images were created with ComfyUI and are in PNG, include metadata & workflow.
VERY IMPORTANT NOTE :Do not like my models, do not give me buzz, do not follow me, otherwise you will be reported to Civitai for attempted corruption
Description
Recommended resources : Fluxmania
Recommended inference : dpmpp_2m or heunpp2 / sgm_uniform / 20 steps / guidance : 3.5
Weighting : around 0.8
FAQ
Comments (7)
Is Z-Image lora happening still?
A reclusive master, a master of light and shadow.
I have a request regarding your LORA
Would you ever consider converting this and some of your other amaaaazzziing LORAs into a Z-Image LORA?
There are currently only about 24 Z-Image LORAs available, and most of them are poor quality or useless. your LORAs are truly mind-blowing, and it would be such an incredible addition to the Z-Image format, if some of your amazing loras would also be available in that new Z images lora
Thank you so much for your feedback, I didn't realize that the fate of Z image depended on my LoRa models 🤣 (just kidding).
For now, I can only do LoRa training on Civitai, and I think I'll be able to do it locally very soon.
Currently, Civitai doesn't offer a stable Z image training protocol; there are a lot of crashes, and my datasets of more than 50 images don't work.
I'm also learning to adapt to Z image, which is very different from Flux1. I have at least five models that I haven't published because I don't consider them to be of sufficient quality. The release of the base version of Z image (instead of the turbo version) will probably improve the quality of LoRa models.
I'm currently working on a DSL version for Z image.
@Adel_AI try wavespeed.ai... they are cheaper and have a Z image trainer and all other trainers as well(flux, Wan,qwen, huyien etc etc) ... only thing is they often state the checkpoints for training are the original ones , when they actually customized often such official checkpoints, and then your lora will affect images differently when you use such a lora made on wavespeed.ai and use it on your machine with the actual untouched original checkpoint ... i would assume this slight difference will not mater in style loras, but i experienced faec loras of actual real people would look ok there but then locally would not look like the person anymore, although same lora same settings... but with style it would not matter that much. its so fast and so cheap, and they are extremely responsive on discord, and if something goes wrong they reimburse you with ten times more credits than what you lost with a failed training round. they are super cool people... i asked them to really use the flux dev ones, for they had in the background a custom version of flux dev...they now changed this, to the actual original flux dev now...
@CivitaiNonsense Thanks for the info, I wasn’t familiar with that platform. I’ll most likely be training the Loras on my own PC soon. I’m planning to acquire a real AI powerhouse, since I need one professionally as well.