NOTE: Civit has made the download UI confusing. For the v4.0 model, it contains four versions:
BF16 UNET (11GB)
FP8 UNET (6GB)
Q6 GGUF UNET (5GB)
FP8 Checkpoint (16GB)
I uploaded them very descriptively but Civit has stripped away the descriptions so it's now confusing to see what you're downloading. The FP8 checkpoint is an all-in-one version that contains the VAE and Text Encoder. For the UNETs you need to provide those extra models.
I've used 100s of curated images to help add some better color, contrast and detail to the Z-Image Turbo model. It's not a huge difference yet, but it is noticeably better in side by side tests. I'm starting out a bit conservative because I don't want to mess up any of the great compositional features that Z-Image offers.
Use the workflow embedded in the showcase images for the model version to get the best results. Check the About this version section for links to the workflow and recommended settings. Z-Image currently mangles nipples and genitalia so my workflow uses SDXL detailers to refine just those aspects of the image.
P.S. I would appreciate any constructive feedback on how to make this model better.
Description
Just trying to make what I hope are small, incremental improvements on the official Z-Image Turbo until the base is released. Use my v1.1 workflow for best results.
FAQ
Comments (24)
Will there be a v1.1 FP8 in the future?
I'm getting a lot of deformed bodies with Pruned Model fp16 v1.1. Legs are getting 6 fingers all the time on 1024x2048 images... It's not my first ZIT model, but first with fp16. Maybe it's not good for my 16 Gb VRAM
Oh interesting. I didn't really do much with the FP16 version. It was a byproduct of the conversion to FP8 so I figured I would offer it as another option. I'll do a bit of testing but I would say to use the BF16 version instead.
Thanks GBRX for contributing to ZIT. This model has huge potentials for how little resources it takes and the quality output it produces. And you are already well known for hot, beautiful images.
Thanks so much!
So your examples are all upscaled with sdxl? Would be nice to see examples from only this model, feels like cheating tbh.
No, that's not right. The showcase images for this model haven't been upscaled by SDXL or even upscaled at all. The workflows are embedded in the images so you should be able to see that the only case where SDXL detailers are being used is to fix the current issue Z-Image has with nude bits. However, these detailers are optional so if you like mangled nipples and genitalia then please feel free to turn them off.
The showcase images that don't have nudity are 100% demonstrating the quality of this model.
@GBRX Well loading up one of your images is how I noticed the sdxl upscale. It was a nude one and I didn't check any others so sorry for that. I still like the model but its still slightly misleading even if its just the nude samples, maybe just mention it since it's not something (any) of the z models are good at atm
@Asaghon Which image created using this checkpoint did you see an SDXL upscale? The only time I used SDXL upscaling was with the official Z-Image model, not this one.
I already mention in the description for this model that the workflow should be used to get the best results and fix the Z-Image anatomical issues. Maybe I can put that in bold or make it more apparent somehow so that nobody else finds it misleading.
Version 1.1 has much better colors and contrast than its predecessors. However, I find the lighting, skin tone, and overall color brilliance inconsistent. Sometimes bright, sometimes dark, sometimes pale, sometimes oversaturated. Also, the face often looks slightly different even with the same prompt. If that's intentional, I understand, but I prefer to adjust settings manually rather than letting the AI do it randomly. (Yes, of course it's not random, but you know what I mean) :)
The face is going to change slightly just because the seed does. That seems to be a characteristic of z-image. Some model architectures maintain consistency across seeds but some don't. I'm not sure about the other issues you're referencing though.
I posted seven pictures here that would have shown what I meant. However, for some reason, they keep getting unpublished.
hm... the pictures are back and published.... however... now you can see, the prompt is nearly the same only clothes but ther are very different
https://civitai.com/images/114299244
https://civitai.com/images/114299238
@Gorean The images are missing all the generation metadata so I have no idea.
@GBRX prompt is open now
@Gorean I tested that prompt and got very similar results both with my model as well as the official z-image model. I did not test it with the skinny lora so I suspect that that lora is having some unwanted effect.
@GBRX It looks like AI has a different color/light profile outdoors and indoors.
@GBRX but i can't explain this:
https://civitai.com/images/114378220
https://civitai.com/images/114378195
@GBRX yours and CyberRealistic Model are the best at the moment :) i like it :) thank you
I checked all the models, this is the only one that changes at least something and does not make the base model worse.
Thank you! That's exactly what I was hoping to achieve - i.e. some kind of improvement without messing anything up.
Many thanks! The best!