Sophos ZiT
I've been playing around with Z-Image Turbo lately. I was throwing together a little of this and a little of that, and I came up with something that felt good enough to share.
This model has an artistic bias for sure. If you like the style, enjoy! If you don't like the style, you probably shouldn't try to shoehorn it into your workflows. See the list below for the models and LoRAs I merged to create this model. You might find something to your liking in there.
📝 Version Notes
Consult the notes below for the differences between the posted versions.
v1.1
This version makes use of some updated components, mostly keeping the same vibe as v1.0 but with improved color saturation and maybe also a sharper image quality.
v1.0
This is the initial version. I think it has a nice aesthetic, is fairly versatile, and seems to hold up well at ~3 MP resolutions.
🤚NOTE: All example images were produced at ~3 MP through my workflow that uses Qwen3.5-4B to enhance the initial prompt. The prompts you'll see on the example images on the CivitAI page were the initial prompts that were enhanced, not the final prompts that were fed to the Sophos ZiT model. For future releases, I will make some adjustments so that my workflow properly saves the final prompt in the metadata.
Don't sleep on LLM prompt enhancement. There's a reason all the cool kids are doing it. Even a small LLM can make a huge difference in image quality if you aren't a prompting virtuoso (or you prefer to be lazy, like me).
Along those lines: I'd say that the majority (~2/3) of the quality in the example images is from the workflow, not specifically this model. I do like how this model handles and I think it brings something to the table that I wasn't finding elsewhere, but I don't want to oversell it. You could take any of the source models I used in the merge, slap the same LoRAs on them, and get mostly the same results with a combination of good prompting + high resolution.
That being said, this model is free and so is my workflow, so what more do you want? 😁 Go make some images and have fun.
Suggested Uses
Realism seems to be what this model is best at doing. I included some example images to show that it can also do animation styles, but I'm sure other ZiT models do animation styles better.
⚠️ NSFW Note ⚠️
This model can do some nudity, but it has all the ZiT shortcomings in the NSFW area. Don't expect exquisite human anatomy below the belt or expect the model to understand kama sutra positions. This is not meant to be "that kind of model." (See my other work if that's more your thing...)
Recommended Workflow
Every example image embeds the ComfyUI workflow I used to generate the image. You can drag and drop any image directly into ComfyUI to load up the workflow.
Three Different Upscale Flows: Face detailer, latent upscale, and SeedVR2. Use all or none. Comparison sliders included.
LLM Prompt Enhancer: This is the secret to getting some really good images. I strongly recommend using an uncensored version of Qwen3.5-4B. The workflow includes a good system prompt and sampler settings. (NOTE: Sometimes the LLM can overwhelm the ZiT model with too many details. It isn't a panacea. Some prompts may perform better if kept simple.)
ClownsharK Samplers: With the supplied sampler settings, I'm getting fairly stable images in the 3 megapixel range. Most of the example images you see were generated without additional upscaling.
NAG Guidance: Ready to go when you need it to stabilize difficult compositions.
Resolution Picker: Requires some initial setup but it is such a godsend once you configure your standard resolutions.
Multi-GPU Support (sort of): Enables the user to load balance different weights across available GPUs, like pinning the model weights to cuda:0 and the text model to cuda:1. It doesn't accelerate the compute at all, but I use it to be able to run fp16 on 2 x RTX 3090s without waiting for weights to swap in and out of RAM. If you only have one GPU, just set all the devices to cuda:0 or swap the multigpu nodes for their single-GPU equivalents.
No Extra Bloat: I tried to keep the workflow as lean as I could while still including useful nodes. Customize to your heart's content!
Merge Details
Credit goes where credit is due. This merge wouldn't have been possible without the efforts of the people who created the following constituent models and LoRAs. Please check out their models, like and subscribe, send them some donations, etc.
NOTE: Someone recommended blending the LoRAs using ComfyUI-LoRA-Optimizer and it does help with quality when using multiple LoRAs with ZiT. v1.0 did not make use of it. v1.1 and beyond are making use of it.
Sophos ZiT v1.0
Checkpoints Used
LoRAs Used
Sophos ZiT v1.1
Checkpoints Used
LoRAs Used
Description
Initial version in fp16 and fp8
FAQ
Comments (13)
The sample images look terrific
Thanks! High res + ralston_2s + beta57 + LLM-enhanced promps = money.
@sophosympatheia What LLM enhanced prompts app do you use? I use LM Studio and its not good.
@Melodic_Possible_582589 My workflow uses llama.cpp locally with a GGUF quant of MuXodious/Qwen3.5-4B-PaperWitch-heresy-v2 (https://huggingface.co/MuXodious/Qwen3.5-4B-PaperWitch-heresy-v2).
I vibe with the style!
"This model has an artistic bias for sure". This website needs more of these. Loved the results.
I've been extensively testing Z-Image checkpoints since they launched, and this one is by far the best at generating realistic, true-to-life pictures of female subjects.
I hope you plan to continue updating this checkpoint. If I may make some suggestions: I feel like the colors could be improved (as they feel desaturated most of the time) and the film grain slightly decreased.
Although both features help add realism, they don't fit well in scenarios that require more vibrancy and detail, and I feel the checkpoint is a bit too heavy on this specific style.
I see you used CyberRealistic Z-Image Turbo v2.0 in your merge, which was also quite desaturated. Check CyberRealistic Z-Image Turbo v3.0 or 4.0 to see what I mean; there is a huge difference in color between them, while the general style remains the same. So tweaking these features won't impact the look or vibe of your checkpoint.
Besides that, GREAT WORK! Thank you very much for putting this together @sophosympatheia
Thanks for those items of feedback! I agree it has some issues with colors right now. I saw that CyberRealistic ZiT v3.0 just became available and I am working on some tests with that currently. I hope to have a v1.1 of the merge out soon.
@sophosympatheia nice to hear that! I do have access to CyberRealistic ZIT v4.0 already, if you want use that on your next merge just let me know. I`m really excited for your v1.1.
Again, great work! Thanks!
Excellent work, I am testing the models one by one and yours is exceptional; it has a unique style but I find it excellent. Well done.
I'm glad you're enjoying it! Thank you for the feedback.