Juggernaut Moss V1

A big thank you to RunDiffusion! They sponsored the compute for this model. Without them Juggernaut Moss would not exist.

Juggernaut Moss is my full finetune of MOSS-Image, made by me, KandooAI, the person behind the Juggernaut XL series. 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.

This is Version 1. It is my first finetune of a brand-new architecture, trained on a fairly small dataset. In my blind tests it clearly beats MOSS-Image Base, but it is not flawless. Hands, small text, large crowds and some art styles can still go wrong now and then. If you find weak spots, please tell me. That feedback goes straight into Version 2.

💼 I'm looking for a job! I'm currently open to new opportunities, either a full-time position or long-term paid contract work in model training, finetuning and image generation. If you like what I build, reach out on LinkedIn or by email to KandooAI@web.de.

Files

Only the transformer was trained. The text encoder, tokenizer, VAE, scheduler and pipeline code in the Diffusers folders are the original ones from MOSS-Image-SFT.

The int8 file was made from the BF16 file. The same quantizer reproduces Comfy-Org's official moss_image_sft_int8_convrot.safetensors bit for bit, so it loads in ComfyUI without any custom node. In a same-seed test it gives practically the same image as BF16. A 4-bit version may follow later.

Recommended settings

These are the settings I recommend:

• Sampler: Euler (flow matching), 30 steps. Anything from 25 to 50 steps works fine. 25 can be a little too low now and then (small details or text go missing), and 50 sometimes overdoes it and changes the composition. Below 25, details start to go missing.
• CFG: 3
• Shift: 5
• Skip layer guidance: layer 22 on the negative branch (the official MOSS default)
• Resolution: around 2 megapixels, for example 1536x1536, 1344x1792, 1792x1344 or 2048x1152. MOSS-Image is at its best between 1.5K and 2K. 1024x1024 works, but looks noticeably weaker.
• Negative prompt 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 prompt for art styles: the same, but without cartoon, cgi, 3d render. Otherwise you are fighting the style you asked for.

How to prompt

Juggernaut Moss was trained on natural-language captions in three lengths: short, medium and long. Short prompts work, but the model really shines with a detailed description. Write it like you are describing a photograph to someone who cannot see it:

1. Subject: who or what, with age, look, clothing and expression.
2. Action and setting: what is happening and where.
3. Light: the most important part for this model. Say where the light comes from and what it does, for example "soft window light from the left", "backlit by a low setting sun", "a single hard spotlight from above, visible beam in the haze" or "neon reflected in wet pavement".
4. Camera and look: lens and aperture ("85mm lens at f/2"), shot type ("close-up", "wide shot") and genre ("editorial portrait photography", "cinematic film still, anamorphic", "documentary street photography").
5. Text in the image: write every word you want to see in quotes, for example a neon sign that reads "Good Vibes Only". Text that is not spelled out in the prompt turns into gibberish.

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.

For skin detail it helps to name it ("visible pores", "freckles", "fine wrinkles", "natural skin texture"). For art styles, name the medium and the technique ("watercolor with wet-on-wet washes and visible paper texture") and leave out camera words.

Usage

ComfyUI: put one of the two .safetensors files from the top level of this repo into ComfyUI/models/diffusion_models/ and load it with Load Diffusion Model. You also need the text encoder and the VAE from Comfy-Org/MOSS-Image:

• text_encoders/qwen3.5_9b_bf16.safetensors, or the smaller qwen3.5_9b_w4a8_gptq.safetensors, goes into models/text_encoders/. Load it with Load CLIP, type moss_image.
• vae/flux2-vae_small.safetensors goes into models/vae/.

Any MOSS-Image workflow works. Just swap in the Juggernaut Moss file.

Diffusers: this follows the official MOSS-Image quick start (pip install "diffusers>=0.38" "transformers>=5.3" "accelerate>=1.10"):

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "Kandoo/Juggernaut-Moss",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)
pipe.to("cuda")

image = pipe(
    "Cinematic film still of a female detective in a beige trench coat standing in the rain on a dark city "
    "rooftop at night, looking out over glowing skyscrapers, her face lit by a red neon sign, anamorphic look, "
    "teal and orange color grade, neo-noir thriller atmosphere.",
    negative_prompt="worst quality, low quality, blurry, jpeg artifacts, watermark, deformed hands, plastic skin",
    height=1152,
    width=2048,
    num_inference_steps=30,
    guidance_scale=3.0,
    uncond_skip_layers=[22],
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("juggernaut_moss.png")

How it was made

• Base: MOSS-Image Base, a full finetune of the transformer (no LoRA).
• Data: more than 20,000 images: my own hand-picked photography and cinematic stills
• Captions: every image has natural-language captions in a short, a medium and a long version, so the model handles both quick and very detailed prompts.
• Training: 1024 px buckets (about 1 megapixel, the same as MOSS-Image's final training stage), learning rate 6e-6 with a cosine schedule, an effective batch size of about 64 images, 6 epochs on a single GPU on RunPod.
• Checkpoint selection: I trained for 6 epochs, rendered a checkpoint at every 0.6 epochs from epoch 3 onward and picked the release in blind tests. The best results came at about 4 to 5 epochs. After that, hands and text started to get worse.

Blind test

Every decision for this model was made in blind tests: random left/right order, no labels, and you pick the better image or "equally good". In the final round, Juggernaut Moss V1 was tested against MOSS-Image Base on new showcase prompts with the same seeds and settings:

Leaving out the ties, Juggernaut Moss won 80 % of the comparisons. It won in almost every category, with the biggest margins in portraits, cinematic, text, street and sports. Landscapes and architecture came out about even.

Known limitations (V1)

• Hands: much better than before, but not perfect. Complex hand poses and hands holding small objects can still fail.
• Text: short words and signs usually work if you spell them out in quotes. Long sentences and very small text are still unreliable.
• Crowds and groups: faces in the background of large groups can get mushy.
• Landscapes and architecture: about the same level as Base. The training focused on people, light and cinematic scenes.
• Art styles: most styles work, but the model leans toward photography. Name the style clearly and use the art-style negative prompt.
• Low resolution: at 1024x1024 the images look clearly weaker than at about 2 MP.

Content notice

The base model can still produce nudity, and Juggernaut Moss has not removed that ability. That is why this repo is tagged not-for-all-audiences. Use the model responsibly and within the law.

Showcase

All images were made with the BF16 model and the settings above (at 50 steps), one seed per prompt, no upscaling and no inpainting. I picked them from a set of 200 showcase renders.

MOSS-Image Base vs. Juggernaut Moss

Same prompt, same seed, same settings. MOSS-Image Base is on the left, Juggernaut Moss on the right.

License

Apache License 2.0, the same as MOSS-Image. You can use the model and its outputs commercially. The bundled text encoder, VAE, tokenizer and pipeline code come from OpenMOSS-Team/MOSS-Image-SFT and are also under Apache 2.0.

Credits

• RunDiffusion for sponsoring the compute.
• The OpenMOSS team for MOSS-Image Base and SFT.
• Comfy-Org for the ComfyUI integration and the int8 ConvRot format, and lodestones (Kromoss) for the open-source ConvRot quantizer.
• KohakuMossTrainer, which this model was trained with.

Contact

KandooAI, here on Hugging Face: huggingface.co/Kandoo

Job offers and collaborations: LinkedIn or KandooAI@web.de