High successful rate, and low style/other parts pollution. Weight=1 works, but can use lower.All above imgs generated with other loras to test whether they can work together, seems great by now. I didn't try realsitic model yet, you can have a try. I trained on anime imgs, so maybe not work.You can do anything you want, no permission needed.======== training suggestion for small items and face experssion =========This is a chanllege problem for me before. Good control on small details (like facial expression) without destroy other content. But finally the answer is quite simple....Use high resolution 1024,1024, then I got the best result so far. (I finally found out that for small items, if under low resolution only several pixel can be seen, how possible model learn correct shape/structure/... from that.)The training recipe is revised from <a target="_blank" rel="ugc" href="http://rentry.org/59xed3">rentry.org/59xed3</a>, but change the epoch number and add alpha(large image need more alpha, lora once fail to learn anything due to low alpha), no large difference for rest parameters. This recipe works really well for all stuff training and saved me a lot of time. Thanks him very much.

rolling eyes and clenched teeth

I am satisfied and tired, I think by now no further version is planned

Files with SHA256: ee0d60b0a61bb73b1329cd8b4e6ef583f179c71d05ddc1cad89481a7077538d1