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
This is actually an offshoot of ZPop v1.0 that I modified to create more Instagram-style images. I've been using it for a while and figured I would publish it in case anyone else could have a use for it. I'm just using a simple workflow that does some extra SDXL refining on certain body parts. You can use the workflow embedded in this image, https://civitai.com/images/124482587.
Tip: I like to prefix my prompts for this model with candid analog amateur jpg.
The Ultralytics bbox files can be found here.
The models used to create this version are ZPop v1.0, Realistic Snapshot, IF-Zimage, and a number of my own private loras based on datasets created with my MoP Mix Big ASP 2.5 and MoP Mix EpicRealism.
FAQ
Comments (35)
Hi, could you show me the documentation on how to properly train a LORA character at your checkpoint?
I don't use my own checkpoints for training z-image loras. I use the defaults in AI Toolkit which seem to work great.
@GBRX If I connect LORA trained on Ziturbo to your Zpop insta checkpoint, the image is spoiled
@zagzagem685 What strength are you using? I tend to keep zit lora strength between 0.40-0.60.
@GBRX At any strength, the image is distorted or the face doesn't look the same. I trained LORA on OneTrainer De turbo 16 GB default settings.
It doesn't seem to play nice with some LoRAs, I'm having similar issues. Without it works fine.
@milt68 The Insta model does not work well with character loras. The new v4 version should work much better.
Did you merge my lora with your finetune?
Yes, the Realistic Snapshot lora is one of the loras I merged in, along with IF-Image and a couple of private dataset loras of mine.
@GBRX Would appreciate some credit or something :)
@MonkeyForever I'm so sorry, you're absolutely right. I had meant to add your lora to the Suggested Resources section prior to release but it slipped my mind. It's there now.
Can you please make gguf Q6_K version checkpoint?
I've had trouble trying to make a GGUF from z-image before but I'll try again.
I've made an open source tool https://github.com/qskousen/ggufy that is easy to use and can create GGUF's of various models, including z-image
@ferretduck Thanks, I'll check it out
@ferretduck I tried your tool, it's very nice. However, I can't get the GGUF for my ZPop model to work correctly. I get a 'Error(s) in loading state_dict for NextDiT' error. If I have time I'll try to dig into it more but do you have any idea what the problem could be?
@GBRX Hmm, i download v30 and it worked fine for me when i converted it to Q8_0. which version of your model did you convert, and what did you convert it to? are you using comfyui-gguf to run it, or something else? i will try to troubleshoot further
@ferretduck I got it to work. I was using the GGUF Loader node instead of the UNET Loader (GGUF) node. Never had that problem before but you learn something new every day ;) Thanks again for the tool!
@45th A Q6_K for v3.0 is uploading now.
"P.P.S. I would appreciate any constructive feedback on how to make this model better."
If your goal was to make a fantastic social network image generator, stop, you've done it.
Ha, thanks! I like it a lot too. I just need to see if I can get better nsfw capability out of it.
why most of these checkpoints are overwhelmed with female photos? is it not capable of creating male figures?
I'm sure it is capable of generating men if that's what you prompt for.
Can you clarify the refiner workflow you recommend to use with this, as in just modify the checkpoint to use this one? Anything else you change?
There are links in the suggested resources section on this page to my z-image refiner workflows. You can also just use the workflows embedded in any of my images.
In your workflow, there's a node that says "nepples" and the numbers in the middle. Mine is glowing red. What is this numerical value? Where does it come from? Do I need to download a file?
You probably need the Ultralytics bbox files, https://civitai.com/models/2172100?modelVersionId=2543959
I put the downloaded files in embeddings. But nodes like Eyes, Pussy, and so on are glowing red. There are no installation errors.
@bionovafood863 No, they go in the /models/ultralytics/bbox folder.
The workflow gives all kind of erros, but using the Checkpoint in other workflows and in Forge Neo is pretty good.
What workflow are you using to get the video output shown here?
My WAN 2.2 I2V video workflow is available at https://ko-fi.com/s/b8848705a0.
Hi could you post FP8 for the latest versions?