This is a slider LoRA. It makes everything less or more spooky depending on the strength. Just in time for Halloween.
Positive values makes everything more comforting, soothing, calming.
Negative values makes everything more spooky, scary, creepy.
Because it's called the spooky slider, it kind of work in reverse, but it just made more sense to me to have the spookiness as a negative and the comforting side as positive. It also work well as a comforting/soothing slider despite its name.
All images in the previews were trained with a very simple prompt and the same seed. I added a comment to the positive prompt of each image with the strength used.
The High noise samples represent the full scale from -12 to 4. The Low noise samples are varied.
Maximum Strength: 4 The video quality degrade when going higher.
Minimum Strength: -12 Video quality drops noticeably if you go lower but that sometimes makes the video creepier, like in some horror movies.
The LoRA was trained with Ostris AI-Toolkit in concept slider mode: https://github.com/ostris/ai-toolkit
The positive training prompt: "comforting, calming, soothing, colorful, joyful, happy, harmless, uplifting, upbeat,"
The negative training prompt: "spooky, eerie, creepy, haunting, terrifying, frightening, sinister, scaring, petrifying, chilling,"
Target class: Atmosphere, environment, characters, objects,
Anchor class: image quality, video quality, lighting,
I anchored lighting because otherwise negatives would just make the scene extremely dark. You can still prompt for darkness of course.
I really like the results at the extremes -10,-11 and +2, +3 but if I had to retrain it, I would remove colorful from the positive training prompt. I think this is why very low values tend become black and white but it fits with the theme.