Style LoRA trained on the artist @18dart3 @18dart5. Please let me know your thoughts!
PonyXL version can also be downloaded here from Huggingface!
Huggingface: https://huggingface.co/AiAF
Twitter: AiArt Factory (@AiArtFactory) / X (twitter.com)
TikTok: https://www.tiktok.com/@ai_art_factory?_t=8cykD6v1PRu&_r=1
Instagram: https://instagram.com/aiart.factory?igshid=NGExMmI2YTkyZg==
Deviantart: https://www.deviantart.com/aiartfactory
Description
D-ART Style LoRA trained off of D-ART's style. I did something a little different with the data set this time. The images used are identical to the ones used to train the [De-watermarked] PonyXL LoRA, I decided to tag them with more natural sounding "naive" tags as opposed to booru style tagging. I also did not include any trigger words for this one. According to this quick-start guide, this leads to the LoRA performing better and has better responsiveness to prompts. It works well most of the time but will occasionally spit out realism pictures. Prompts like "anime style picture" or "cartoon" help steer it back to D-ART's style.
FAQ
Comments (4)
Very nice! Are you training flux style loras using the recommended settings, 20-30 images at 512 squares?
The data set used for the flux version is nearly identical to the one used to train the De-watermarked Pony version (You can download that data set here). The only difference is that instead of using booru style tagging (many comma separated words), I used a clip based tiger. Supposedly more natural sounding tagging results and better performance so I gave that a try. I did not drop any of the images either. I just retagged them and tossed The data set into the trainer.
peak fiction
Cool