RABBITH0LE is a LoRA trained for internet horror, analog photos, cursed found-footage atmosphere, wet darkness, harsh flash, uncanny figures and creepypasta-style visual texture.

This is not a clean cinematic horror LoRA, and definitely not gothic pretty-horror. I wanted it to feel more like something found in an old forum thread, a corrupted archive, a police evidence folder, a trail-camera dump, or a deleted imageboard post.

No trigger word required. Just use descriptive horror / photo language in the prompt.

Recommended strength:
0.6 - 1.1 for balanced results
1.2 - 1.5 for more aggressive output

Higher strengths can get darker, dirtier, more distorted, and less clean. Sometimes that is exactly the point, but if the image starts falling apart too much, lower the strength a bit. There are other examples below.

Note:

This was honestly a difficult LoRA to train. The style I wanted was already dirty, degraded, low-quality, and unstable by nature. That is what gives it the cursed internet-horror feeling, but it also made the training process harder. If pushed too far, the model could easily become broken instead of just creepy.

The final version came from training the raw dirty style first, then continuing from the best checkpoint with a softer refinement pass and improved captions. That gave me the best balance between corrupted analog texture, dark atmosphere, and creepy figures / objects.

Training path: first stage at lr 0.000155, linear 128 / conv 64, then refined at lr 0.000065 with caption dropout 0.18.