Style LoRA trained on on NoobAI-XL vpred 1.0 with images from the primarily NSFW artist D-ART:
https://twitter.com/18dart5
Trigger word is jz235
It works on mentioned model and its mixes:
https://civarchive.com/models/833294/noobai-xl-nai-xl
Negatives and quality tags are optional. CFG around 3-4 is enough.
v2.0 is new and more faithful to D-Arts's style and in general seems to have better anatomy. All previous watermarks that appeared in the corners of images are no longer there (i removed them from the training images. I recommend v2.0 as the model of choice.
v1.0 is smoother and more clean in looks but is underfit and has a hard time with applying it's style on characters that are overfit on NoobAI (like Asuka (soryuu asuka langley)) and has worse anatomy.
Description
MUCH larger dataset (85 images) - first autotagged and then manually checked, removing unnecessary tags/adding missing tags to images as needed.
ALL watermarks removed from training images (yes, ALL. Thanks IOpaint!)
pre-downscaled images before training using custom image downscaling script
sd-scripts does this already, but I wanted to achieve a higher downscale quality, using a script that downscales the images using a combination of Lanczos (40 %) and LSID (60%) algorithm by blending images from both algorithms together at mentioned percentages.
LSID from paper: https://arxiv.org/abs/2510.24334
using edm instead of minsnr for timestep correction/training
final lora model is a 50/50 blend of epoch 25 and epoch 35, using the lora-merger in sd-scripts.
resized lora using this script, with a threshold of 2.85, shaving off 30-35 MBs and reducing overfitting:
FAQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.