Juggernaut XL Ragnarok
A big thank you to RunDiffusion - Juggernaut would not exist without them.
Juggernaut XL Ragnarok is the last SDXL Juggernaut, made by KandooAI. It was released on CivitAI back then, but never on Hugging Face. This is the official upload by its creator.
It is an SDXL 1.0 checkpoint for photoreal people, portraits, cinematic scenes and a wide range of art styles.
Content notice: This model can create nudity and other adult content. It is tagged
not-for-all-audiences. Use it responsibly, in line with the law and the license's use restrictions.
Files
| File | Format | Size | Use |
|---|---|---|---|
juggernautXL_ragnarok.safetensors |
FP16 single file (VAE in FP32), original release | 7.1 GB | ComfyUI, Forge, A1111, SD.Next |
juggernautXL_ragnarok_fp8_e4m3fn.safetensors |
FP8 single file: UNet in float8_e4m3fn, text encoders FP16, VAE FP32 | 4.2 GB | ComfyUI, Forge (less disk space and VRAM) |
model_index.json, unet/, text_encoder/, text_encoder_2/, tokenizer/, tokenizer_2/, vae/, scheduler/ |
Diffusers, FP16 | 6.5 GB | 🧨 diffusers |
The FP16 single file is the original release file (SHA256 dd08fa32f98d05a2443ca1419e46df1575a0811f6e3b246d9dd47ff20f5eb66a). The FP8 and Diffusers versions were converted from it, nothing was retrained or merged. In a same-seed test the FP8 file keeps the composition and the level of detail of the FP16 file; only small details differ.
Recommended settings
These are the settings used for every showcase image on this page. They are read from the PNG metadata, which the images in images/ still carry, so you can drop them into Forge or A1111.
- Sampler: DPM++ 2M SDE, scheduler Karras
- Steps: 30
- CFG: 5 for portraits, 6 for everything else
- Resolution: 896x1152 (other SDXL sizes such as 832x1216 or 1024x1024 work too)
- Negative prompt: none was used for the showcase. If you want one:
bad eyes, cgi, airbrushed, plastic, deformed, watermark - Hires fix: 1.5x with 4x_NMKD-Siax_200k, 15 steps, denoise 0.33, hires CFG 4
- VAE: the showcase used sdxl-vae-fp16-fix. The baked-in VAE works too. If you run it in FP16 and get black or broken images, switch to the fp16-fix VAE.
All prompts with seeds are in showcase_prompts.txt.
Usage
ComfyUI / Forge / A1111: put one of the single files into your checkpoints folder (models/checkpoints in ComfyUI, models/Stable-diffusion in Forge and A1111).
Diffusers:
import torch
from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
pipe = StableDiffusionXLPipeline.from_pretrained(
"Kandoo/Juggernaut-XL-Ragnarok", torch_dtype=torch.float16, use_safetensors=True
).to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
pipe.scheduler.config, algorithm_type="sde-dpmsolver++", use_karras_sigmas=True
)
image = pipe(
"photograph close up portrait of a female boxer training, serious, stoic, cinematic, dark moody",
num_inference_steps=30, guidance_scale=5.0, width=896, height=1152,
).images[0]
image.save("ragnarok.png")
Showcase
|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
License
CreativeML OpenRAIL-M. You can use the model and its outputs commercially. You may not use it for the purposes listed in the license's use restrictions. If you share the model or a derivative, you must pass the same use restrictions on.
Contact
KandooAI, here on Hugging Face: huggingface.co/Kandoo