v1.7Well, this took me a bit longer than expected ... (long story short : I hate python scripts).This new version was trained on a far larger dataset, and this time I manually removed the colored screen from all visual novel samples to have a consistent black screen as in the anime version. As a result, there is less variability in the result and no strange colors artifacts/details.I also manually edited some pictures to create wide angle shots, precisely tagged camera angles, positions and field of view. In my tests, this allows more diverse views compared to the first version which mostly worked on portrait and close-ups.I also changed the trigger, for this one use Eprchkr. Sometimes when the character is typically wearing a collar, you will end up with a mix of the collar and the euphoria choker. In this case, I found that adding something like "black choker" or "jewelry" in negative prompt give a better result. You can also try "white Eprchkr" or increase the trigger weight, but the improvement is usually moderate.I feel this one still have lot of room for improvement, maybe I will try to improve it later.Feel free to share your creations !______________________v0.1I recently assembled a large dataset of Euphoria pictures, and after training the Andou Miyako model, I noticed that the collar didn't render so well. So I figured why not try a clothing/object Lora on this aspect.It turned out far more difficult than I expected, but I learned a lot about training parameters in the process.This version should be considered alpha, with a lot of issues which need solving. Unfortunately I won't have time to do it in the near future, but I plan to work on this when I'll have some time.In particular generation of wide shots leads to low resolution of the collar. There's also occasional conflicts when a character wears a collar. Finally, the green and red lights on the collar randomly get mixed up.The activation words are "euphoria collar". Sometimes "white euphoria collar" helps. Stronger weights, up to "euphoria collar:1.2" might also help.Trained on IllustriousXL0.1, tested on WAI-Illustrious-SDXL 14.0

Euphoria Collar

Huge change of training set and parameters.Lesser impact on style and position of subjects, especially better results for "full body" prompts.

Files with SHA256: 6bb3383114ed4193594bc44d5532e15983bb71f42d64d52a9ad118bf53da097d