For business inquires, commercial licensing, custom models, and consultation contact me under [email protected]
Use Juggernaut on HuggingFace (Safetensors and Diffuser Files)
Update:
Juggernaut-Z Fast is online. The model is excellent for rapid image generation to find the right composition. Afterwards, I recommend a second pass with a different model (e.g., Juggernaut Z, Z-Image Turbo, etc.)
Recommended Settings for Fast are:
Steps: 4-8
CFG: 1-1.5
Sampler: DDIM (there are a lot more Samplers that are working good with it )
Hello everyone,
It has been almost exactly one year since the last Juggernaut version (Ragnarok). After some downtime, TongyLab and BlackForest Lab provided us with two excellent community models (Z-Image, FluxKlein) at the end of last year.
I am presenting the first output of this here: Juggernaut Z, a fine-tune of the Z-Image base.
The checkpoints are available in all standard formats (bf16, fp8, and various GGUF versions). If I missed a specific format you need, just let me know in the comments. Also, a quick heads-up: if you already have the Z-Image Base running on your setup, there is no need to download the Text Encoder and VAE again, you can simply reuse your existing ones
I will keep this brief. This is version 1, so naturally, not everything is perfect yet. It will take 2-3 versions to fully integrate everything we have planned. Please be patient, as work is simultaneously underway on the Juggernaut variants for FluxKlein 4B and 9B.
Our main focus for this first release was killing the glossy "AI plastic" look of the base model. You will notice much better micro-contrasts, actual skin pores, and physically accurate lighting. We also improved the structural logic to reduce hallucinated geometry.
A Quick Word on Prompting: Please actually read the Prompt Guide. If you stick to the old SDXL tag-style prompting (just throwing comma-separated keywords at it), you might get a lucky shot here and there, but you will mostly run into errors and broken concepts. Juggernaut Z thrives on semantic prompting, talk to it in natural, descriptive sentences. If you are not used to this yet, I highly recommend using your favorite LLM (GPT, Gemini, Claude, etc.) to help structure your ideas into proper sentences.
I have included a ComfyUI workflow, which should ensure problem-free generation.
I recommend following Settings with the Workflow:
Resolution: 960*1440 or something similiar with that Pixelsize. "Low Res like 1024*1024" will sometime looks to grainy/noisy
First pass
Steps: 22
Denoise: 1.00
Sampler: Res_2s
Scheduler: Beta
Second pass
Steps: 3
Denoise: 0.15
Sampler: Res_2s
Scheduler: Normal
The foundational workflow was created by nsfwVariant. It was only minimally modified for our purposes, so credit goes to him. Please show him some support.
Also you can find a couple of Comparison Shots in my Profile and also on the RunDiffusion Juggernaut Z Page mentioned above :)
Any LoRA trained on the Z-Image Base is fully compatible with Juggernaut Z. At least, every single one I’ve tested so far worked flawlessly.
Finally, have fun generating. We will be back soon with a FluxKlein fine-tune and Version 2 of Juggernaut Z.
Description
CFG: 1-1.5
Steps: 4-8
FAQ
Comments (24)
Thank you so much for the fast version of the model! I think it will be in high demand among users with limited VRAM. As for image quality, I wouldn't say it's significantly worse; you can always fine-tune the quality with LoRAs, and there's certainly no shortage of those available.
My first constantly used model a Juggernaut-sdxl and Z-image base with Juggernaut give huge advantages. Perfect for creating textures
Good afternoon!Please tell me if it is possible to use the RunDiffision model in ForgeNeo? And if there is such an opportunity, can you tell me which settings to use? I can't set it up in any way.....The output is a terrible image....
I actually don´t use Forge, so i cant really help. But after the Release of Juggernaut-Z other People had similiar Problems with ForgeNeo and got it fixed. Maybe this helps for you:
https://civitai.red/models/2600510/juggernaut-z?dialog=commentThread&commentId=1186534
@KandooAI Thanks! I'll take a look.
@Dev_Nomad I'm using it in Forge Neo. My settings are in my images metadata. :)
@The_Folie Good evening! Thanks, I'll take a look!
with the fast version I found that loras seem to break the image generation with wild distortions as if the cfg was turned too high. turning the loras off the image generation is fine. turbo base works fine with the loras in my workflow. any idea why that might happen? i'm sure it is user error because others in the comments say loras work fine.
Are u using Z-Image Turbo LoRa´s or Z-Image Base LoRa´s ?
So what's with license? Here it said Apache but on HF it's NC — non-commercial use only
Hello, ComfyUI has just been updated and now supports the native use of Int8 ConvRot versions. Do you plan to release an Int8 ConvRot version of this model (which I'm a fan of—thanks! :P)? Int8 ConvRot is faster and offers better quality than FP8 (yes, even on RTX 4XXX/5XXX cards).
Hey! Thanks for the heads-up. I just baked and uploaded the native INT8-ConvRot versions for both Juggernaut-Z and Juggernaut-Z Fast here on CivitAI.
Since CivitAI doesn't have a dedicated selection tag for 'INT8' yet, I had to list them under 'FP8' for now. I'll change the format tag as soon as they add official support for it.
Also, both versions will be uploaded to Hugging Face later today as well.
@KandooAI Incredible responsiveness—thanks for making the Int8 convrot version available; I’m going to test it right away. In my initial tests on Z image base (Normal version), we’re seeing a speed boost of around 27% compared to FP8, along with slightly better quality—it’s a win-win. :)
I tested the available INT8 Convrot version; unfortunately, there seems to be a problem—it only produces blurred images: https://imgur.com/a/zMbU9Pm
Z image base INT8 convrot (normal ZiB version) works without problem but not the Juggernaut Z INT 8 CONVROT for some reasons.
There might have been a problem with the conversion.
@Illynir Prob my fault, i didnt test it. Gonna work on it today, and once i confirm it works i will upload it again. Looks indeed that there was a problem with the conversion
@Illynir Hey, I sat down and worked on the Juggernaut-Z Base version. The conversion took a little while, but I finally found the solution, tested the model, and everything works beautifully. I've just added this version under the "Optional Files" section.
I will follow up with the Juggernaut-Z Fast version over the weekend; unfortunately, I ran out of time today.
@KandooAI Thanks, i will test.
By the way, Siveroxides created a very simple and fast script to convert any model to int8 here: https://github.com/silveroxides/convert_to_quant
This is the script Comfy uses to create the official models they have already uploaded to their Hugging Face page.
It only works with a BF16 model version, for obvious reasons.
EDIT: Works perfectly and speed is on par with ZiB INT8 Convrot (normal version). Thanks. :)
@Illynir I did indeed use the Silveroxides Script, but still took a while (like 90-120 minutes). Anyway, since it works i am fine with it :)
Jugernaut-Z Fast Int8 will be uploaded on the weekend :)
Hi everyone,
Quick update on Juggernaut-Z V2: The model will drop by the end of this week or early next week. Right now, I’m in the final stages of testing and deciding which exact checkpoint will be the final release version.
To give you a little teaser, I just uploaded a few test images from V2 right here: https://civitai.red/posts/29773135 Just a heads-up: these are all raw One-Shots. No cherry-picking involved, so bugs and flaws are included ;)
At the same time, I want to officially announce that my partnership with RunDiffusion has expired. I am officially a free agent now, but don’t worry, Juggernaut will absolutely continue. Also a Huge Thanks to RunDiffusion for the Last 2 1/2 Years. It was an awesome ride :)
Being independent has its pros and cons. The biggest downside is definitely that running high-end training and infrastructure is simply not sustainable with just my own private resources. I just can't finance that out of my own pocket. Because of this, I am actively looking for new sponsors or platform partnerships to keep this thing running long-term. If you represent a platform or know someone, feel free to reach out and of course, the community is highly welcome to recommend me to others :P
The massive upside, however, is absolute freedom. For the past few years, I was barely allowed to touch the NSFW department due to corporate restrictions, which in my opinion, really hurt the model's overall capabilities. That freedom is finally back, and I think you will already notice the improvements in V2, or at the latest, in the upcoming versions :).
Thanks for standing by me, and get ready for V2!
Can your model generate explicit pornographic sex images?
No
how to use the page to train Checkpoint models, Stable Diffusion or Flux models, Z image turbo, and much, much more.