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 the same as v1.1 FP8 except that it can work as a normal checkpoint. So it includes the text encoder and the UltraFlux_V1 VAE because it provides much sharper details than the normal Flux VAE.
I haven't created and published a separate workflow for this but you can simply drag one of the showcase images into ComfyUI to get a usable workflow.
FAQ
Comments (24)
Looks amazing. I am looking forward to having some time to test it !
I wouldn't have used the smokes in the examples, but they are great nonetheless !
add images as pure z-image coz sdxl refiner images not called as z-image ,like stable yogi and zmania
I'm not sure what you're referring to with stable yogi but the showcase images for the new AIO version of this model are 100% z-image.
some one answer pls, dos this work on forge?
Looks like Neo Forge v2.6 maybe, https://github.com/Haoming02/sd-webui-forge-classic/releases/tag/2.6
why r u still using forge. use comfy
@KK_Biswal it's bad , takes too much time, and too complicated ,
@bigdog11 If ComfyUI was "bad, takes too much time and was too complicated" then it wouldn't be the #1 tool for professionals. It's perfectly fine to just say you don't like it and don't want to invest the time to learn how to use it properly.
Did that 2.6 version of Neo Forge work for you?
@bigdog11 its only bad, takes too much time, and too comlicated bc you havent learned it. it allows for MUCH better images simply due to its "complexity" because your limited by what forge "lets" you do. whereas comfy lets you do literally anything lol. skill issue
hahaha guys you speak as if it as rocket science, I used comfy for some time, I'm a professional with nodes in massive engines not this simple stuff, but it's not for me , I like simple ui , it's just stupid ai so |I'm not gonna invest time into it, I make better results with forge
@bigdog11 WE speak of it as if its rocket science??? YOUR the one that said "its bad, takes too much time, and too complictated" and im just saying its only bad and complicated because you havent learned it. if you were actually a "proffesional with nodes" youd be using comfy. theres literally 0 benefit to using anything else in comparision to comfy but hey, whatever floats your boat ig.
Try SwarmUI, it has simple interface like forge but if you want it runs comfyui in backend. Forge is just too outdated now. Comfyui even got lots of templates and community workflows. You don't even need to invest that much time to learn it. And it got the fastest support for any new released model. Follow their huggingface page.
The models looking really good, congrats!. Seen as you asked for constructive feedback to make the model better, I can give you some ideas as I am in a similar situation with some of my trains, trying to get that last 10% of finalising texture and detail.
A trick I used to use with SDXL, and it works well with Z-Image to get those details and "pop" without altering the base model or any training thats been done, is to make a low-rank texture lora to use or blend on top. This works because you limit the amount of control the lora has to effect other elements like geometry and composition, by using a low rank.. and by feeding it only texture and pop, thats what it learns.
So you take your original dataset, and create a texture and pop version, cropping closeups of everything, and then train with a setup of something like rank 4-8, LR 3e-5.. decreasing to 1e5 for 20-40 Epochs, at 1536.. (Z-Image was trained at 1500 so anything over that is training ontop of upscaled noise). Maybe this is something you do or have done before, but I thought I'd share.. You can do this in a short blast at the end of training, but then you don't have the flexibility to experiment, to see what blend works best.
Thanks for the suggestions. I'll look into implementing this.
Can you post an image with the embedded workflow and I'll take a look?
It's an Ultralytics python-package issue. You have to downgrade it. I was having zero success as well but it got fixed after the downgrade. FYI, the detector is just not that good. Forget about detecting anything even through sheer fishnet or in bad lighting, even with ridiculously low threshold values.
Works fine for me. There seem to be two main ultralytics BBOX nipple models floating around out there: nipple_yolov8s.pt (about 22 mb) and nipple.pt (35 mb). GBRX has one named Nipples.pt in his workflows, which I suspect is the 35 mb file renamed, as that one, for me, gives consistently better results than the yolov8s.
@ViceCityMojo I didn't find that 35MB one.
@cecilialennyy552 https://civitai.com/models/135658/nipple-adetailer-model
@GBRX If I train a character's lore on ZImage with good nipples, will ZImage have detailed nipples?
@zagzagem685 You'll probably get an improvement but I doubt they'll be SDXL quality.
I've uploaded the bbox files that I use. https://civitai.com/models/2172100?modelVersionId=2543959
isnt nipple detailer made for the original zimage loader? ive noticed incompatability with certain loras using different custom checkpoints for each model. some results good, some bad and play with the weight of the lora models.