Join The Tinkerer on Whop to download this model and many more — all for one monthly payment. Plus, get early releases, private tools and a lot more. 👉 Join on Whop
CyberRealistic Z-Image Turbo is a build based on the original Z-Image Turbo. It follows the original Z-Image Turbo settings and philosophy, with minimal intervention, exploring how well this setup translates into the CyberRealistic workflow.
⚙️ Personal Settings (Forge Neo)
or
Required Additional Files
Make sure you also have the following:
16 GB+ VRAM: qwen3_4b.safetensors
8–12 GB VRAM: qwen34bfp8_scaled.safetensors
All VRAM sizes VAE: ae.safetensors
This model is shared as-is, for users who enjoy experimenting, benchmarking, and pushing models outside the comfort zone.
Feedback, findings, and edge cases are welcome - this release exists primarily to learn from real-world usage.
Description
FAQ
Comments (61)
what 'trained' means ? is this trained !? post xy compare images here please
@amazingbeauty You are not allowed to write down this question! I've asked the same what he really did with this checkpoint and he was hiding my comment! It's just download farming and stats pushing cause of the z-image hype. Too bad!
You will get the same results like the original model. Nothing changed here.
@MrSmith2025 hide your comment? How? (OK I noticed it was hidden. Sorry)
Some compare images: https://civitai.com/posts/25098314
are you using any upscalers, etc? The images look quite sharp and clean even for the medium shot lengths.
I hear this more often. Personally, I use a modified version of Forge Neo, and it gives good results. ComfyUI is great - but I just don’t have consistently positive experiences with it myself.
@Cyberdelia hi, i use forge neo also. mind sharing what was modified or recommend an extension or upscaler? I've used quite a few of the popular upscalers and have used resharpen, but it doesn't really make a difference because z image's output is just as good. I did hear that the higher vram version output is way better. I am using the lower vram version, but have also seen people have clean and clear results with some extra things.
@Melodic_Possible_582589 you know ReSharpen? That will help a lot
@Melodic_Possible_582589 I will try to publish some of my custom extensions.
I havent had time to test it myself, but from the example pics Its kinda lost its unique ZIT realism that Ive grown to love, going back to a slightly uncanny synthetic look.
After I test it out with my own settings I will delete this criticism if Im wrong.
That's a big problem in general with all the merged and trained models of ZIT distilled, they all seem to devolve back to the SDXL look. I dont know if everyone is training on flux or SDXL data or its just because its a distilled model.
This is not a finetune, and distilled models are not worth to fully finetune. And no sane person trains on synthetic data, its just not working the same way as in a normal model. Flux finetunes are the same, not a real advancement(except Chroma).
@sonnychimaobi439 You’re partly right. This really should be seen as an experiment - there is training involved, but it only truly becomes interesting once the full base model is released
@Cyberdelia Yes. And we will lose the speed. Anyway do you still use Neo? With ny 8gb card immediataly OOM, except if i enable never OOM, but very slow.
@sonnychimaobi439 Yes, still Forge Neo. It’s been modified in such a way that I honestly don’t feel the need to switch yet. It works fine, though 8 GB can be a bit tight for Z-Image.
@Cyberdelia I will try on 1.5 resolution(512x768), maybe its not that bad.
At least the model's creator is being honest and isn't trying to coast on the popularity of his previous models (the CyberRealistic Pony is my favorite among all SDXL family realistic models) and sell this one for 5000 buzz.
@WhateverName It would be easy for Cyberdelia to put all models in one page, and be number1 in all category forever:)
@sonnychimaobi439 That’s the peak. After that, I’m basically living on borrowed time
1. Awesome, thank you! 🙏
2. Where is Cyberrealistic Chroma?? 😭
A guy can dream.
Yeah we need Chroma !!
@Bonticarius Thank you for your supporting my influence campaign. ;)
It’s great news that you’ve started working with Z-Image Turbo.
This feels like a very promising direction.
Considering how strong your SDXL and Pony models already are, it’s exciting to imagine what CyberRealistic could become on ZIT.
I really hope this will grow into one of the most popular realism models in the near future.
Looking forward to seeing how it evolves 👍
Do I have to download and use vae and encoders along with this model? Sorry I'm new to Z-Image.
Please drop links to dependent files too like you did with the Flux. Thanks!
Yes, you download them here https://civitai.com/models/2168935/z-image?modelVersionId=2442439
@Cyberdelia thanks
VAE is the same as Flux btw
I’ve added the files to the description. The filenames are different, but the contents are the same as in the link I provided.
Thanks Cyberdelia.
Need a FP8 6GB GUFF version to try it, if no problem.
I have a GTX 1070 8GB VRAM.
I have the same GPU and this version can be run with no issue :)
@MalkhaiC how long take it to generate an image?
@Cyberdelia 13 minutes with upscale and face detailer. A Flux model takes about 1 hour to do the same with my GPU 😅
@MalkhaiC damn!
yeah you should be fine. im running fp16 on a 6gb RTX 3050 (75w) and offloading to 32gb system memory , using swarmui at 13 steps at 1216x832 with 1.5x upscale ( 4 steps) it takes about 4 minutes per generation. i can do face upscaling or swapping using reactor and it might add a few seconds to the time, shoutout too the budget build brigade lol 😀
@MalkhaiC DAMN!
how do i run z-image model i have RTX 4060 8gb vram but still it crashes
@Cyberdelia depending on configuration. I started to try zimage 1 week ago.
First impresions on Q8:
Quality: worse than Flux
Speed: better than Flux.
Tomorrow i wll post here the logs. Resolution/steps
p.d: my test were with [ZIT] Z-Image Turbo Kijai (fp8_Scaled_e4m3fn)
@frikitin Quality: “worse than Flux” usually indicates an issue with your configuration. Under normal circumstances, the quality should be better, although that can be somewhat subjective. Also, are you using fp8_Scaled_e4m3fn instead of qwen_3_4b?
@Cyberdelia im testing other models Pruned Models fp8 (5.73 GB) for GTX1000 series.
This model no.
for text encoder i use qwen_3_4b
@Wendy_Earth @forfreelsd368 try:
--windows-standalone-build --listen --force-fp16 --lowvram --preview-method auto
@frikitin thanks, but mine DAMN was because I got same timings of creating pics on rtx2070 years ago.
@MalkhaiC THANKS!! for the comment. It works with my gtx1070!!
But i dont understand why. With FLUX any model over 8GB dont work unless be in GUFF format. ¿?
@Cyberdelia After testing i think its a problem with the sampler used.
A lot of people recomends euler-simple with z-image models ¿?
The best sampler for this model is the recomended in the instructions.
dpm++2s_ancestral - beta (In comfyui v.0.4 there is not DPM++ 2s a RF ¿? what is RF??)
Using that sampler the quality gains a lot. The other ones are grainy and blurred, specially in backgrounds.
NVIDIA GTX 1070
PROMPT1 832x1216
4steps ..... 336seg euler - simple FIRST TIME
4steps ..... 110seg euler - simple SECOND
10steps ..... 253seg euler - simple
PROMPT2 832x1216
10steps ..... 225seg euler - simple
10steps ..... 348seg dpm++2s_ancestral - beta
10steps ..... 174seg res_multistep - simple
14steps ..... 344seg euler - sgm_uniform
PROMPT3 1024x1328
14steps ..... 460seg euler - sgm_uniform
14steps ..... 628seg dpm++2s_ancestral - beta
PROMPT4 1024x1328
14steps ..... 628seg dpm++2s_ancestral - beta
Love this. Just a point of info, a lot of people advocate a Euler A / DDIM set-up with ZIT but that combo creates very weird outputs from this model.
Edit: It was something to dowith Aura Flow and EulerA - probably just my bad
Really? And this is not a problem with the standard version?
@Cyberdelia No the standard version likes EulerA/DDIM. But the tone of the images I get are brighter and the prompt adherence is messy and the body is often confused.
It's not a major issue for me - I'm getting really great images with this model (thanks again).
EulerA is working for me with SGM_Uniform and simple etc. It might just be the combo or DDIM.
I'll do some more tests and upload later :)
@Cyberdelia It's something to do with AuraFlow when I turn it off and user Eulera/DDimuniform it's fine. Probably my issue
@Zeddy456 i was testing different schedulers Does DDIM scheduler goes really well with euler A?? Also what auraflow you are using? im always confused about auraflow if it needs to be activated, changed or not?
@brahianvalles I use Aura Flow 3 for everything else. And yeah for me DDIM Uniform and EulerA works well (but not with Aura Flow) ymmv
res_multistep with simple schedular works pretty good too.
At first: THANK YOU! For your great Work.
But this one doesnt work for me.
I get this error message
: CLIPSetLastLayer
'NoneType' object has no attribute 'clone'
I know.. its in early state, but maybe this helps to find bugs.
I'm your biggest fan!
Samples look gorgeous.
I know its gonna take a lot of time and training but... Can you consider to detroy the "flux chin" (intrinsic to Zit) in future installments? thanks!
Love you, man. Don’t tell the others.
I don't know if ZIT has a Flux Chin issue, only when it if finetuned on Flux images, that is my experience anyway.
@J1B It’s also almost not present. And when it is, it just looks natural. We shouldn’t label every chin as a Flux chin.
@Cyberdelia really? ok, maybe it was an impression i got from looking at some samples. I could be wrong. Have a nice day plox.
@ElectricDreams It could also just be that it’s no longer noticeable by me anymore :)
@Cyberdelia Yeah about 10%-15% of the population have a cleft chin IRL (including myself) and seemingly about 50% of Hollywood actors: https://www.buzzfeed.com/mjs538/celebs-with-cleft-chins
So I am also not sure why people scream "Flux Chin!" at the first hint of them.
@Cyberdelia Please upload a Pruned fp8 model
It's uploaded
@Cyberdelia How long does it take you to fine-tune Z-Image? and How much does it usually cost?
Finetune would be hard. Easier to just do LoRA or Lycoris.
how do I convert it to GGUF?