A krea2 bf16 rank 128 style LoRA for making Danbooru 1024 explicit and questionable images.

Trained on a 40K subset of my Danbooru 1024 dataset https://huggingface.co/datasets/quarterturn/danbooru-1024-eq-captioned. The dataset was further filtered to exclude any images tagged as having censorship or English text.

The dataset was character tagged in conjunction with the Danbooru-2026-clean tag database, feeding artist and subject information into the captioning process for each image.

The resulting lora should assist krea2 with genning anime characters by name only, or emulating an artist style, though results will vary based on how well the artist or subject is represented in the subset of the dataset used for training.

Donations

Donations in any amount are gratefully accepted to support future datasets and research costs.

Ethereum (ETH): 0x73a3fB9EB1fa259b7395F8cBBa3C1025453530d6

Downloads last month
421
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for quarterturn/krea2-danbooru-1024-40k

Base model

krea/Krea-2-Raw
Adapter
(616)
this model