This is a modified version of z-image turbo.
After some try a found a way to modify model, to pushing out details and do clear images, especially on background and microdetails, where original model tend to grain (see comparison image).
The modification is balanced to not destroy text part.
Credit's to:
original z-image turbo creator
me, for modification.
If you use this model for training, merge please respect my work and give back link.
Uploaded to have more test and feddback. Thank you for who try it and give feedback.
Description
v1.0khv
FAQ
Comments (21)
lora merge?
No. The model is original, with my own method to modify it. Anything is merged into the model.
@devilkkw tnx coz all merges on civit brake loras. Gonna test it
@alexsmileface thank you so much
Are there any plans to release an f16 version?
Duo my limited vram, actually no. Sorry
@devilkkw What's about RAM and a huge swap file?
@qek you mean something like offload?
@devilkkw Offloading
@qek i've to try if work.
@devilkkw This is really good work. I'm happy to put my 1tb ram + 96gb vram RTX6000 to work on training you a fp16 model if you are willing to provide instructions and your training data for how to do it. I don't want any credit, I just would like the full fat quality that's possible since Z.ai hasn't released the base model yet.
I realize you may want to keep the secret sauce to yourself, so no worries if you say no. Just offering.
@q5sys well you could buy @devilkkw cloud credits and it protects his work
@svenhimmelvarg That would cost quite a bit more money than just letting my machine run at full tilt and paying the 0.065kW/h power rate I have. I'd happily sign an NDA, I do that for most of my paying clients, so I'm used to it.
For context, I've been an open source dev for over two decades now, I'm the Fedora Release Maintainer for four Fedora Releases (Security Lab, Games, Jam, Astronomy). My github and linkedin is the same screen name that I have here, so anyone can verify that I'm not a random shady character online. haha
I agree this method seems to be good results. Maybe 20% better than base model. I like that this is not a merge - because most other models right now lose quality/detail due to some merging or using other model's outputs. Can you suggest sampler settings / workflow for this? Also would really like if you can describe your method . Thank you for your work!
For the setting and sampler, I suggest to download the image and drag it in comfyUi.
If you don't use comfyUI, sempler are er_sde with simple scheduler, cfg 1, 8 step.
Also, for some trick you can read this https://www.reddit.com/r/StableDiffusion/s/Sba1VukXC8 where I posted some trick.
Also, my method is based on old method I use for sd1.5, working on layer, understanding it, and pushing out what i need.
I just posted the model for feedback, to see if it works with many different settings people use.
v1.0 i like the checkpoint :) good color, good contrast, nice understanding and this for an fp8. Thank you for it :)
Thank you so much for feedback.
The differences are so subtle that it's hard to tell exactly what's improved. What technology was used to improve the details?
How do we get the full bf16 model?
can we get this in int8 convrot ?