This is my personal workflow that I wasn't planning on uploading anywhere, but I decided to share it with anyone who's interested. It's designed to "pump out the maximum possible quality and details" from the model we all love so much ;)
Resources:
https://huggingface.co/ostris/Z-Image-De-Turbo/tree/main
or
https://huggingface.co/leejet/Z-Image-Turbo-GGUF/tree/main
or
NSFW > https://huggingface.co/tewea/z_image_turbo_bf16_nsfw/tree/main
***
https://civarchive.com/models/2193783/z-image-uncensored-text-encoder-abliterated-huihui-qwen3-4b-v2-q8-gguf
https://huggingface.co/Kim2091/UltraSharpV2/tree/main
(optional LORA): https://huggingface.co/wcde/Z-Image-Turbo-DeJPEG-Lora/blob/main/dejpeg_v2.safetensors
flux gguf vae from: https://huggingface.co/calcuis/pig-vae/tree/main
Description
if you do not like it - use your own ;)
FAQ
Comments (27)
I din't try it but it seems to be made for impatient little pervs who can't wait a tuned NSFW base model
:)
Olivia is a trend setter ;)
@MetaGen TY!
I tried it, but I'm not impressed, its a fancy workflow, but my janky and simple UltimateSDupscale workflow can give much better quality results. I used the exact same parameters as the metadata but got terrible results, maybe I'm doing it wrong. Can you upload your raw Json? it usually has the verbatim workflow while the image metadata does not.
I suspect all the quality boost is coming from the lora you linked and not the overly complicated workflow. The lora alone does a good job increasing the quality of an image on its own on the most simple workflow with one basic ksampler and SDupscaler.
I'm too busy rn to support my WF, sorry
how can I adapt this workflow to produce larger images, whenever i change the empty latent to like 1080p or more i get crazy pixelization toward the bottom of the image only. Also, what does the optional lora do?
+ node - SDImageUpscale
8mn for 1 gen with default settings on my RTX4080.
Quality is good but I can get almost as good quality in half the time with SDXL refiner :/
almost as good with SDXL is not about hands ;)
@0l1v1aR0551 wouldn't it be better to use union controlnet to inpaint face + hands at low steps before sdxl refiner? i could imagine 8 steps base latent + 6 steps controlnet inpainting + 24 steps sdxl img2img at like 0.17 denoising strength
@MetaGen just let SDXL go to its deserved cozy grave already ;)
@0l1v1aR0551 Sorry but if I wanted to gen 1 pic every 10mn I would use Chroma not a turbo model.
@MetaGen Chroma is not as good as ZIT
@0l1v1aR0551 distilled is always worse for creative artists.
@MetaGen yes, but here is the secret - ZIT is not distilled ;)
try to use negative prompt on it - and you will be amazed (+ more steps than 8)
@0l1v1aR0551 I tried negative prompt in my first ZIT workflow, it had almost no impact. it might not be distilled but it acts as distilled, at least the Q6_K version I am using.
@MetaGen distillation do not depends from quantization, you probably were just doing few steps, instead you have to add serious negative prompt and do 20-40 steps to see the big difference
@0l1v1aR0551 Have you actually tried to use a negative prompt? every time I do it seems to reinforce it like a second positive prompt instead.
@MetaGen it is used this workflow at cfg=2.5 ;)
@0l1v1aR0551 https://civitai.com/models/2196015?modelVersionId=2472641
ZIT is indeed distilled. People use the de-distilled model for lora training.
@MetaGen I know that authors say that it is "distilled" and that there is a de-distilled version by Ostris (for LORA training) ... but, the very fact that this model is VERY sensitive to CFG > 1 = means - it is a mix of distilled model and the original one, so you can do generation in 8 steps with cfg = 1 and also do it with 20-40 steps with cfg = 2.5 - 4.5 and the result will be way different and better
@0l1v1aR0551 I get pretty good results on 9 steps and it takes only 1.5mn with Q6_K for 1k+ res
Also increasing the cfg scale messes with lora quality, but yes it helps with negative prompt. on 1cfg it barely acknowledges 1.5 weights at start of negative clip. i generated 200+ images with explicit negative prompt and got maybe 10 obeying.
@MetaGen cfg scale is optimal at 2.5, LORAs mostly love simple euler when we push their limits in this ZIT model
Doesn't work with lora.
no - it works with lora