Version 2.0 recommended! This version is so much better - it is more stable and consistent, and faithful to the original. With this change, Dr. Midou is now out of the prototype phase.You can compare the showcase of the two versions to see the striking differences.Start with weight of 1.0, but you can go to 0.9 / 0.8 if you feel she looks stiff and overfitting.So what has changed in this version?I have made many LoRA training attempts in the past two months, and a ton of lessons learned. Here are the most important ones:<ul><li>Use a low network dim! Flux is a very powerful yet delicate base model. It has been tuned very well to produce high quality images, and your LoRA can totally ruin the fine balance. At higher ranks, it is more often to see bad anatomy, bad hands, and distorted faces. Do yourself a favor and lower it to 2 - it will learn characters just fine.</li><li>Pay attention to your training data! Unlike Stable Diffusion which is much less sensitive and you can get by with a small portion of bad images, Flux will attempt to learn every single image you put in. Images with complex composition / tricky pose will really confuse your training. For example, I removed all pictures like these from Emiko's training set - you don't want Flux to draw this orgasm expression out of blue. Also, twisted body pose training image will lead to bad anatomy.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f6ba01ff-ef52-415c-b2b2-13afa517eea3/width=525/f6ba01ff-ef52-415c-b2b2-13afa517eea3.jpeg" /></li><li>Pay attention to the consistency to your training images! If you have images that are supposed to be of the same person but drastically different, Flux will be confused and not able to converge. For Emiko's case, her art style changed considerably from Innai Kansen 1 to the sequels. I removed all the Innai Kansen 1 version (the image on the top) and the model consistency improved greatly.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1ffb23fe-2af7-4ecb-ada1-c4865752f336/width=525/1ffb23fe-2af7-4ecb-ada1-c4865752f336.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/342a969c-a543-40bb-8252-5937e8167c7d/width=525/342a969c-a543-40bb-8252-5937e8167c7d.jpeg" /></li><li>Training resolution doesn't matter very much. In Emiko's case, I used 512x512 since the images themselves aren't high resolution. The same rule as Stable Diffusion training: NEVER use AI to hires your training images!</li><li>Finally, I saw no effect on natural language captions. The danbooru tags generated by WD14 work just fine in Flux. Follow the same rules as Stable Diffusion training: remove any tag that you want to be intrinsic to the model. For Emiko's case, I removed any tag for her hair length, hair and eye color.</li></ul>I hope this helps!

Innai Kansen - Midou Emiko

A retrained model with refined dataset and lower network dim of 2.

Files with SHA256: b8352ce3fcbada16e4694cc4a2893246e21de7f22ae86987dd85271565c3368a