Gives Z-Image the Jib Mix treatment, it can force more realism from fantasy prompts as well as improving NSFW anatomy.
Make the default faces more European and less likely Asian.
If you want less soft/blurry image outputs, use the Ultra Flux 4K VAE instead of the standard VAE.
I am currently mainly using it around 0.20 - 0.60 strength on model and clip, as it is a little strong and makes Z-Image look like Flux if used too strong.
It was tuned to work with my ClownsharkSampler setting and it doesn't seem to play that nicely with the onsite generation that maxes out at 15 steps and a low resolution.
09/12/2025 - Fixed auto-generated file name. (was 7BKEAXV6TGEH07FD2D20FYRJPO.safetensors)
11/12/2025 - Updated Workflow to V2.
Added a powerful seed variation custom node to make images much more diverse: https://github.com/ChangeTheConstants/SeedVarianceEnhancer
Added Power Lora Loader now it is supported by Z-Image.
Description
Initial release, can make images look "Fluxy" at higher values.
FAQ
Comments (16)
Did you use ai-toolkit, or diffusion-pipe, or something else?
I've tried ostris' training adapter and de-distilled model, and loras trained on the de-distilled model (same dataset/settings) have come out much better (less loss of fidelity compared to without loras), but still harm the quality a fair bit (though not enough to be unusable, until ~10k steps when training on the de-distilled model, vs around 2/3rds of that with the adapter). https://github.com/None9527/None_Z-image-Turbo_trainer seems promising - almost done with a comparative run of it now.
Running some prompts from your qwen model through to see how it compares - hard to match that qwen realism in 6B, but z-image does appear to produce a much more consistent image, with brighter colors that only the SDXL upscale would get me with the qwen model.
This one was trained on Civitai as I was feeling a bit lazy, but I have just kicked of a V2 training run on AI Toolkit with Ostris De-distilled model so we will see how that turns out! :)
Qwen is still really good, but it has its issues which are mostly solved when using Z-image.
Nice, I was waiting for this. Thanks!
Another Z Lora turning realistic Z faces and emotions into same looking classic AI look faces with default smile.
Yes, Z-Image base model is very hard to beat as it is almost totally believable, but it does wander off prompt into a fantasy art style with Asian faces and too much noise at times, with some prompts, which this lora can help with.
Perfecto!
Amazing, Unbelievable !!!!
where the png metadata? need to get the workflow from it :;3
Some of my images have metadate attached: https://civitai.green/images/113584245
If you look up I have also posted the workflow: https://civitai.green/models/2194714?modelVersionId=2481800
I have to say that despite Z image base model isn't even out yet. This is a very incredible LORa. Really helps elevate the already impressive Z image turbo. Do you recommend ComfyUI-Detail- Daemon too? Also overall do you think this is better than Qwen image especially your Jib mix Qwen v5?
I think as a whole package it is better than my Qwen, mainly because it is faster but also I was getting a lot of issues with big bobble-heads and short legs with my Qwen model that z-image doesn't have. My suggested Workflow is just to the right on the top bar, using ClownsharkSampler and tuneable noise options. I will upload a V3 workflow later as I have found the Ultra Flux VAE better. It's a good start for Z-Image but I think we can push it a lot further when the full none Turbo model comes out to train on.
I think everyone is overtraining with Z-Image Turbo. Since the results are already strongly visible with 8 epochs and 400 steps (also with only 4 images used it gives good style output). I been using 400 steps for styles and 1200 steps for characters models and can maintain 1.0 stength with good results. The upside is: It takes less time to create a Lora now. (I'm using musubi-tuner locally for ZIT Lora's within ComfyUI) 400 steps takes me 28 minutes , and 1200 steps approxly 1 hour on 16GB VRAM... The amount of images used doesn't really affect the time much, so It's just amazing. Only minor issue mixing multilple Lora's is a bit buggy with Z-Image..
I interesting, yes you may well be right, I will have to try that out.
I did try testing my early lora Epochs around 500 steps at strength 1 but I didn't see much improvement, they still looked a lot better at lower wieghts. Groking often doesn't occur until around 7000 steps: https://m.youtube.com/watch?v=D8GOeCFFby4
Sounds super low, over 90 loras ive trained 1200 steps would get OK results but at around 4400 steps the detail likeness and flexibility is miles better. Also with 1536 x 1536 -
Though I am about to try Musubi in Comfyui Realtime Lora for next training, so would be great if this turns out to be the case!
Details
Files
Jibs_Realistic_z-image_lora_V1.safetensors
Mirrors
145_Jibs_Realistic_z-image_dec18.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Jib_Mix_Realistic V1.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Jibs_Realistic_z-image_lora_V1.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.