logs:

• v1 is unreliable use v2;

• target name:

  • v1:

    • "^.*output.blocks.1.*attn.*$", for unet;

  • v2:

    • "^.*output.blocks.1.*attn.*$", for unet;

    • "^.*(q|v).proj.*$", for text enc;

• size:

  • v1 (~5mb);

  • v2 (~8mb);

• trigger:

  • v1 no trigger

  • v2 3d-filter

• lora str:

  • v2:

    • 0.5-1 (between 0.5 - 0.8, is giving semi 3d/real);

• use hiresfix with 0.4-0.6 denoise will clean up some artifacts (especially fingers (see workflow img)) and keep cfg on the low end (~3 - 5);

• sampler: euler a;

additional:

• v2 is trained with 24 img from slash-soft and 1 other 3d artist, I forgot the name. Compared to v1 (which is trained on 500ish img) is giving much better result. I think because with v2 I manually tag the image while v1 is using wd tagger and not using trigger? or maybe because v2 has better hyperparams and text enc is trained to?

• Maybe I will upload/train a third version with more img and using real life cosplay photo;

• v2 face has a bias towards tifa and aerith (because that's ~90% the datasets lol).