A big thank you to RunDiffusion! They sponsored the compute for Juggernaut Moss, the model this Turbo is built on. Without them none of this would exist.
Juggernaut Moss Turbo is Juggernaut Moss with my own DMD2 turbo LoRA baked in. It makes images in 6 steps at CFG 1 instead of 30 steps at CFG 3. On my RTX A4500 a 1344x1792 image takes about 11 seconds instead of more than a minute. I didn't want to use one of the existing speed LoRAs. I wanted one made specifically for Juggernaut Moss, so I trained it myself. If you'd rather load it as a LoRA on top of Juggernaut Moss, you can find it here: Juggernaut Moss DMD2 LoRA (Beta).
This is a Beta. The results are close to the full model, but the Turbo has its own character: it picks safer compositions and gives less variety between seeds. The next version will be trained on BF16 and I'm already working on improvements. If you find weak spots, please tell me.
💼 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.
| File | Format | Size | Use |
|---|---|---|---|
juggernaut_moss_turbo_v1_int8_convrot.safetensors |
int8 ConvRot (W8A8), same format as Comfy-Org's moss_image_*_int8_convrot files |
5.4 GB | ComfyUI, recommended |
juggernaut_moss_turbo_v1_bf16.safetensors |
BF16 single file (ComfyUI layout) | 10.2 GB | ComfyUI, experimental (see below) |
juggernaut_moss_turbo_v1_fp8_scaled.safetensors |
FP8 (E4M3, scaled) | 5.4 GB | ComfyUI, very close to BF16, works on any card |
juggernaut_moss_turbo_v1_w6a8.safetensors |
W6A8 (6-bit weights, ConvRot) | 4.5 GB | ComfyUI, smallest file that stays very close to BF16 |
juggernaut_moss_turbo_v1_nvfp4.safetensors |
NVFP4 | 3.3 GB | ComfyUI, for RTX 50 / Blackwell, see the note below |
juggernaut_moss_turbo_v1-Q8_0.gguf |
GGUF Q8_0 | 5.7 GB | ComfyUI with the MOSS GGUF nodes, practically the same as BF16 |
juggernaut_moss_turbo_v1-Q4_0.gguf |
GGUF Q4_0 | 3.3 GB | ComfyUI with the MOSS GGUF nodes, for very low VRAM, changes the image more |
model_index.json, transformer/, text_encoder/, tokenizer/, vae/, scheduler/, pipeline_moss_image.py |
Diffusers | ~28 GB | 🧨 diffusers |
workflows/Juggernaut_Moss_Turbo_int8.json |
ComfyUI workflow | drag it into ComfyUI |
int8 or BF16? The turbo LoRA was trained on the int8 version of Juggernaut Moss, and the int8 Turbo is the version I tested the most. In my tests it was also the better one: the BF16 Turbo made more mistakes (an extra limb here, a missing hand there) and is slower. So for this Beta, int8 is the recommendation. The BF16 file is there for everyone who needs it, and the next version will be trained on BF16 so both are equally good.
The smaller versions (fp8, W6A8, NVFP4 and the two GGUFs) were added after the release. I render-tested every file with the same prompts and seeds against BF16. fp8, W6A8 and Q8_0 come out practically the same as BF16, Q4_0 and NVFP4 change the image noticeably more. A 4-bit ConvRot version didn't make the cut, it simply looked too bad. The safetensors files load with Load Diffusion Model in an up-to-date ComfyUI (I tested with ComfyUI from mid-November 2026, older versions may not know the newer formats yet).
A note on NVFP4: NVFP4 is made for RTX 50 / Blackwell cards, which run it natively and fast. On older cards ComfyUI has to unpack it on the fly, which works but is slow, so there fp8 or W6A8 is the better choice. It is also clearly the weakest of the smaller files: compositions drift further away from BF16, and fine textures can get a bit grainy, especially in dark scenes. Only use it if you really need the small file.
Only the transformer is new. The text encoder, tokenizer, VAE, scheduler and pipeline code in the Diffusers folders are the original ones from MOSS-Image-SFT, the same as in Juggernaut Moss.
Juggernaut Moss was trained on detailed natural-language captions, and the Turbo likes the same kind of prompt. The structure that works best follows the training captions:
Text in the image: write every word you want to see in quotes, for example a neon sign that reads "OPEN ALL NIGHT". And this matters even more with the Turbo: if there is a place where text would naturally appear (a sign, a label, a jersey, a book cover), prompt that text too. Otherwise the model fills it with gibberish. Telling it "no other text" does not work.
Example:
A cinematic film still of a street musician in his forties playing the saxophone alone under a stone
bridge at night, eyes closed, completely lost in the music. He wears a long dark coat and a knitted grey
hat, and an open saxophone case at his feet holds a few coins and a handwritten cardboard sign reading
"THANK YOU". The shot is a medium shot taken from a low angle in three-quarter profile. The lighting is
low-key, with warm orange light from a street lamp on one side of his face and cool blue light from the
river on the other. The image has a fine film grain and soft halation around the light sources. Shot
with a Sony ILCE-7M3, 35mm lens at f/1.4.
ComfyUI: put juggernaut_moss_turbo_v1_int8_convrot.safetensors into ComfyUI/models/diffusion_models/ and load it with Load Diffusion Model, or just drag workflows/Juggernaut_Moss_Turbo_int8.json into ComfyUI. You also need the text encoder and the VAE from Comfy-Org/MOSS-Image:
text_encoders/qwen3.5_9b_int8_convrot.safetensors (or the bf16 / w4a8_gptq version) goes into models/text_encoders/. Load it with Load CLIP, type moss_image.vae/flux2-vae_small.safetensors goes into models/vae/.GGUF: the GGUF files use the MOSS layout of the community MOSS GGUF nodes. You need city96/ComfyUI-GGUF, the official OpenMOSS ComfyUI nodes and RealRebelAI/ComfyUI-MOSS-GGUF. Load the file with MOSS-Image GGUF / BF16 Transformer (Rebel), the text encoder with MOSS-Image Qwen3.5 Encoder (Rebel) and the VAE with MOSS-Image VAE (Rebel), and put a ModelSamplingSD3 node with shift 5 between the model and the sampler. Without it the images come out dark and washed out. Everything else stays the same. Keep the Turbo settings: 6 steps, CFG 1, euler_ancestral / simple.
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-Turbo-Beta",
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(
"A 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 that "
"reads \"HOTEL\". The shot is a medium shot taken from a low angle. The image has a fine film grain.",
height=1792,
width=1344,
num_inference_steps=6,
guidance_scale=1.0,
generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("juggernaut_moss_turbo.png")
MOSS-Image can produce nudity, and neither Juggernaut Moss nor the Turbo removes that ability. That is why this repo is tagged not-for-all-audiences. Use it responsibly and within the law.
All images were made with the int8 Turbo, 6 steps, CFG 1, euler_ancestral / simple, one seed per prompt, no upscaling and no inpainting. I picked them from several hundred renders. The prompts and seeds are in images/showcase/PROMPTS.txt.
Apache License 2.0, the same as Juggernaut Moss and 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.
KandooAI, here on Hugging Face: huggingface.co/Kandoo
Job offers and collaborations: LinkedIn or KandooAI@web.de
juggernaut_moss_turbo_v1_bf16.safetensors 10 GBjuggernaut_moss_turbo_v1_int8_convrot.safetensors 5.4 GBjuggernaut_moss_turbo_v1_fp8_scaled.safetensors 5.4 GBtransformer/diffusion_pytorch_model-00001-of-00003.safetensors 5.0 GBtext_encoder/model-00002-of-00004.safetensors 5.0 GBtransformer/diffusion_pytorch_model-00002-of-00003.safetensors 5.0 GBtext_encoder/model-00001-of-00004.safetensors 5.0 GBtext_encoder/model-00003-of-00004.safetensors 5.0 GBjuggernaut_moss_turbo_v1_w6a8.safetensors 4.5 GBjuggernaut_moss_turbo_v1_nvfp4.safetensors 3.3 GBtext_encoder/model-00004-of-00004.safetensors 1.9 GBvae/diffusion_pytorch_model.safetensors 250 MBtransformer/diffusion_pytorch_model-00003-of-00003.safetensors 190 MBKandoo/Juggernaut-Moss