A Journey Into the Macabre

Step into the dark and mysterious world of Victorian Gothic Horror, where shadows linger, and the line between human and inhuman blurs. This LoRA model is inspired by the chilling elegance of 19th-century gothic literature and art, blending the haunting atmosphere of Victorian horror with a focus on the uncanny, the grotesque, and the supernatural.

Designed to generate images in a sepia-toned, aged aesthetic, the model evokes the feeling of old, faded photographs or once elegant paintings that have witnessed too many horrors. It brings to life not-quite-human subjects, from eerie figures cloaked in fog to strange creatures lurking in the shadows of decaying mansions. The unsettling yet beautiful results capture the eerie stillness and creeping dread that define the gothic horror genre.

Whether you're looking to create haunting portraits, eerie landscapes, or chilling supernatural scenes, this LoRA model will provide the tools to explore the dark, atmospheric world of Victorian Gothic Horror.

This LoRA is made as part of the KI Welten Discord Group Collaboration. For the other Great Halloween-themed Flux LoRAs, follow the link to the KiWelten Halloween Collection.

Usage

To use the most recent version of the LoRA, use the following settings:

• Trigger word: vicgoth

• Other tokens that work well: Mist, fog, skeletal, shadows, mask, cloak, and forest.

• Lora Strength: Full Strength (1.0) is recommended, but play with ranges from 0.8 to 1.2

For differences in previous versions, see the version notes below.

Version Notes

Version 5.0

Version 5.0 was improved by editing the dataset (fixing the demon horses) and enhancing the captioning of the images so that the LoRA embodies much of the tone and emotion of the dataset's images as well as the style. With the new captions, the LoRA is better at producing dark, sepia tones and adding disturbing and unsettling qualities without additional prompting. The framing of subjects has also been improved. Version 5.0 is the first to be trained on multiple resolutions, hopefully adding better prompt adherence.

Version 4.0

I added to the previous version's dataset to include an expanded selection of horror imagery for Version 4.0. The result is a more robust style that includes the tones and art mediums that I wanted to capture from the training dataset. The LoRA no longer needs to be prompted with specific keywords like sepia or photograph, and it will now apply an appropriate style fitting the descriptions in the prompt. I've found it works when you let the model choose by not specifying any style. Put the trigger word "vicgoth" at the start of your prompt and use the LoRA at a strength between 0.6 and 1.0. Version 4.0 was also an experiment with a more efficient training technique (a larger batch size and additional gradient accumulation steps with a scaled learning rate), which turned out well. You can read more about these settings in the Training Settings Related to Datasets section of my Dataset Preparation Guide.

Note: Due to an error in building the dataset, all horses in Version 4.0 are demon horses. You will have nightmares.

Versions 2.0 and 3.0

Version 1.0 broke Civitai's Flux LoRA trainer, so I had to train version 2.0 on my hardware. Version 3.0 was trained with improved captioning to improve the style when applied. To get the full effect, use it at a strength of 1.0 and the trigger word "vicgoth" along with "sepia-toned photograph." For those that tried v2.0, you'll be happy that v3.0 follows prompts better, doesn't add horns to everyone just because, and deepens the creep factor significantly.