--v4.2--

This is a vast improvement from version 2. I made v2 mostly for inpainting purposes, which suited my purposes well enough. It was very iffy with full character generation - a lot of artifacting and more often than not generated poor results when used for that purpose.

I used the same dataset as version 2, which uses HQ images, but changed the captioning approach and used an updated workflow and the results were much better.

This version fixes that and now will is fairly consistent with realistic high quality character generations, and of course can still be used for inpainting.

--v2--

Culled some of the previous images, added a few more to get more consistent uncut outputs. I changed the captioning with a different strategy so that in this version you shouldn't need to use a trigger word to get the best results.

While I intended to make this one more of an inpainter but it works for full character generations fairly well too.

--v1--

This LoRA was generated from a dataset and generational input that @Gaydevai_paulo provided - I only donated GPU time to it generate, so if you like it please give him the props.

I uploaded what I thought was the best version out of this 10K step run. My testing shows that this is generally good at up-close shots, but you may need a few generations for images that are full medium shots or wider. I found that if the overall composition is what you wanted then it makes a great inpainter.