stylish LoRA (PonyXL)版本V3 (ID: 668299)

Testing loras on civitai. Models might be hit or miss since I can't test before publishing.

I created this model to have a single LoRA that has a style similar to these images: image 1, image 2 (NSFW), image 3 (NSFW).

Model was training using generated images from a combination of the following:

• SD model - Pony XL

• LoRa (style) - Suoiresnu

• LoRa (style) - ExpressiveH

• LoRa (concept) - Blackedwear

• LoRa (style) - Bartolomeobari

• LoRa (style) - Sinfully Stylish

• LoRa (style) - fishine

• LoRa (slider) - Detail Slider

Changelog:

• v1 - the model is workable, but not really what I wanted (see training data)

• v2 - added more images and tags, switched to Adafactor

  • simple images might need halo in the negative

• v3 - added a lot more variation to the images

• v3.1 - same as v3, but trained with Prodigy and less repeats/steps - this one is my favourite

• v4 (in progress) - will try to reduce overfitting

描述:

Training: defaults (Adafactor), 412 auto tagged images (20 tags, 0.4 threshold)

训练词语: