Essenz – Darkest Dungeon [Chris Bourassa] (Style LoRa for SDXL 1.0)

An SDXL 1.0 LoRa made by AI_Characters for producing images with an artstyle similar to the game Darkest Dungeon – with the art mainly created by Chris Bourassa.

Usage recommendations/settings used for the sample images:

Weight: 1.0 (may need to be adjusted down when used in conjunction with other LoRa’s)

Model: SDXL 1.0 with 0.9 vae

Generation parameters: 25 steps | DPM++ 2M Karras | CFG 7

Positive prompt: high-quality (the model was trained with this tag)

Negative prompt: watermark, text, signature, blur, low-resolution, low-quality

CLIPTextEncodeSDXL ComfyUI node resolution: 4x your latent image resolution (so 4096×4096 for a 1024×1024 image) (yes it affects output and makes it slightly better)

Trigger words

• “artwork in dgst artstyle”

Changelog

v1.0

• Initial release

Further notes

It doesn’t replicate the style perfectly as unfortunately it does not produce the typical black/not visible eyes as seen ingame. I tried a few different things to no avail. Adding “eyeless” in front of a person prompt may help a little bit.

Additional Info

Please credit me when using this model in a merge or hosting it on a generation service.

I share all of my model’s and training knowledge for free, but donations to my Ko-Fi would be greatly appreciated!

Training

Dataset – 76x artworks| Config |

Workflow:

1. Train full finetune using the above config on a Vast.AI rented A6000 GPU with 48gb VRAM

2. Find somewhat overtrained epoch (but not too overtrained)

3. Extract LoRa using these settings

4. Resize LoRa using these settings