A huge thank you to RunDiffusion for sponsoring the compute time for this model. Without them Juggernaut Moss wouldn't exist!

I'm currently looking for a job! Full-time or long-term contract work in model training, finetuning and image generation. If you like what I do, reach out on LinkedIn or send me a mail at [email protected] :)

Juggernaut Moss on HuggingFace: https://huggingface.co/Kandoo/Juggernaut-Moss

Hello everyone,

After Ragnarok on SDXL and Juggernaut Z on Z-Image Base, here comes the next one: Juggernaut Moss V1, my full finetune of MOSS-Image Base. :)

I tested MOSS-Image a lot over the last few weeks and was honestly impressed by how good it already looks out of the box. So of course I wanted to give it the Juggernaut treatment.

The goal was the same as always with Juggernaut: photography and cinematic images that look real, with natural skin, believable light and a modern film look, while the model can still do everything else.

But please keep in mind: this is Version 1. It's my first finetune on MOSS, and the dataset was fairly small. It's not flawless. Hands, small text, big crowds and some art styles can still go wrong now and then. If you find weak spots, let me know in the comments. All of that goes straight into Version 2.

I made every decision for this model in blind tests (random order, no labels, just "which one is better?"). In the final round against MOSS-Image Base, with the same prompts, seeds and settings, Juggernaut Moss won 73 % of all comparisons. Base won 18 %, and 9 % were equally good. Leave out the ties and that's an 80 % win rate, with the biggest gaps in portraits and cinematic shots. Try it yourself and tell me what you think. :)

I put some Comparison Shots and Step Tests below in the Gallery. So take a look :)

Which file should I take?

BF16 is the full-quality version (~10 GB). The int8 ConvRot version is half the size (~5.4 GB) and faster on most GPUs. It uses the same format as Comfy-Org's official MOSS int8 files, so it loads in ComfyUI without any custom nodes, and with the same seed it gives you practically the same image as BF16. If you want to use Diffusers, you'll find that version on HuggingFace.

How to use it in ComfyUI

Put the file into ComfyUI/models/diffusion_models/ and load it with "Load Diffusion Model". The text encoder and VAE are not baked in. You need the ones from Comfy-Org/MOSS-Image (https://huggingface.co/Comfy-Org/MOSS-Image):
Text Encoder: qwen3.5_9b_bf16.safetensors (or the smaller qwen3.5_9b_w4a8_gptq) in models/text_encoders/, load it with "Load CLIP", type moss_image
VAE: flux2-vae_small.safetensors in models/vae/

If you already have MOSS-Image running on your setup, there's no need to download the Text Encoder and VAE again, you can simply reuse your existing ones. After that, any MOSS-Image workflow works. Just swap in Juggernaut Moss.

Recommended Settings:
Res: around 2 MP, e.g. 1344*1792 for portrait, 1792*1344 for landscape, 1536*1536 or 2048*1152 (MOSS really shines between 1.5K and 2K, 1024*1024 looks noticeably weaker)
Sampler: Euler
Steps: 30 (anything between 25 and 50 works, 25 can be a bit low sometimes and 50 sometimes overdoes it)
CFG: 3
Shift: 5
Skip Layer Guidance: Layer 22 (the MOSS default)
Negative for photos: worst quality, low quality, blurry, jpeg artifacts, watermark, signature, text artifacts, deformed hands, extra fingers, bad anatomy, distorted face, plastic skin, oversaturated, cartoon, cgi, 3d render
Negative for art styles: the same, but without "cartoon, cgi, 3d render", otherwise you're fighting the style you're asking for


Prompting

Juggernaut Moss was trained with natural language captions in short, medium and long versions. Short prompts work, but the model really comes alive when you describe the image like you're describing a photo to someone who can't see it: who or what, what's happening and where, the light, and the camera.

Light is the most important part. Tell the model where it comes from and what it does, like "soft window light from the left", "backlit by a low setting sun" or "neon reflected in wet pavement". Lens and shot type help too ("85mm lens at f/2", "close-up", "cinematic film still, anamorphic"). If you want skin detail, just name it: "visible pores", "freckles", "natural skin texture".

If you're still used to the old SDXL tag style (just throwing comma-separated keywords at it), you'll get a lucky shot here and there, but mostly worse results. If writing full sentences isn't your thing yet, let your favorite LLM (GPT, Gemini, Claude, etc.) turn your idea into a proper description.

Text in the image: write every word you want to see in quotes, e.g. a neon sign that reads "Good Vibes Only". Text that isn't spelled out in the prompt turns into gibberish.

For art styles, name the medium and the technique (e.g. "watercolor with wet-on-wet washes and visible paper texture") and leave the camera words out.

Example:
Close-up portrait of a 34-year-old Nigerian woman with deep brown skin, natural coily hair pulled back, small gold hoop earrings, looking straight into the lens with a calm, confident expression. Soft window light from the left, a gentle falloff into shadow on the right side of her face, visible pores, fine vellus hair on her cheek and a subtle natural sheen on her forehead. Neutral warm-grey background, 85mm lens at f/2, shallow depth of field, high-end editorial portrait photography.

What to keep in mind (V1):
Hands are much better, but complex poses and hands holding small objects can still fail. Short words and signs work well, long sentences and tiny text are still hit or miss. Faces in the background of big crowds can get mushy. Landscapes and architecture are about on the level of Base, because the training focused on people, light and cinematic scenes. And the model leans towards photography, so name art styles clearly.

A few words about the training:
Full finetune of MOSS-Image Base, no LoRA. More than 20,000 images: my own hand-picked photography and cinematic stills. Every image has a short, medium and long caption. Trained at 1024px for 6 epochs, and the final version was picked in blind tests. Epoch 4 to 5 looked best, after that hands and text started to get worse.

And in case you're wondering: Juggernaut Moss is Apache 2.0, just like MOSS-Image. You can do anything you want with it: merge it, train on it, sell your image outputs, etc. Just a simple shoutout is all I ask. :)

Have fun trying it out. As always, I'm eagerly waiting for your pictures in the Gallery! :)
If you liked the model, please leave a Like. In the end, that's what helps me the most as a creator on CivitAI.

Last but not least, I'd like to thank a few people and projects without whom Juggernaut Moss wouldn't have been possible:

RunDiffusion (for sponsoring the compute)
The OpenMOSS Team (for MOSS-Image)
Comfy-Org (ComfyUI integration and the int8 ConvRot format) and the Kromoss project (open-source quantizer)
KohakuMossTrainer