logs:<ul><li>v1 is unreliable use v2;</li><li>target name:<ul><li>v1:<ul><li>"^.*output.blocks.1.*attn.*$", for unet;</li></ul></li><li>v2:<ul><li>"^.*output.blocks.1.*attn.*$", for unet;</li><li>"^.*(q|v).proj.*$", for text enc;</li></ul></li></ul></li><li>size:<ul><li>v1 (~5mb);</li><li>v2 (~8mb);</li></ul></li><li>trigger:<ul><li>v1 no trigger</li><li>v2 3d-filter</li></ul></li><li>lora str:<ul><li>v2:<ul><li>0.5-1 (between 0.5 - 0.8, is giving semi 3d/real);</li></ul></li></ul></li><li>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);</li><li>sampler: euler a;</li></ul>additional:<ul><li>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?</li><li>Maybe I will upload/train a third version with more img and using real life cosplay photo;</li><li>v2 face has a bias towards tifa and aerith (because that's ~90% the datasets lol).</li></ul>

Files with SHA256: 0b44d88337251d4cec56ae02a5527c14150c26005c73882cf61682484e2439ee