Trained from 50 curated images from the artist jibble_jobble. If you're a fan of Jib's work, now you can generate that same aesthetic yourself, more consistently than the original, and completely free.

Why Kroma

This LoRA was trained for Kroma, and Kroma is currently the best uncensored base model around. It's a Krea 2 fine-tune that was purpose built for NSFW. The big thing you'll notice is prompt adherence. Kroma actually follows what you tell it, even with complex scenes and a lot of detail. No fighting the model, no garbled nonsense, no refusal to do what it was literally trained for. It just works.

You're getting Jib's style on a much better foundation than the old SDXL workflows.

Better than the original

JibbleJobble uses an older SDXL base model with a stack of different LoRAs on top. That means their output drifts a lot between generations. Different LoRAs fight each other and the style gets inconsistent.

This is a single LoRA, one style, weight 1. You get the look more reliably and at higher quality, even with complex prompts that would turn into soup on Jib's setup.

How to prompt

There's no trigger word. Weight should stay at 1.0. Use Euler Simple.

Don't dump a wall of Danbooru tags. This model responds best to natural language. Describe what you want like you'd describe a scene to someone. Structured prompts work too if that's your thing, YAML or JSON style.

Example:

> A young woman with pastel pink hair, soft smile, looking directly at the viewer. Warm golden rim lighting catching the edges of her silhouette. The background is a glowing cityscape at night, blurred into soft bokeh. The

> atmosphere is dreamy and intimate. Masterpiece quality.