This has more image to train, so it looks much better.More preview on my Pixiv: <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/18763851">4BEraser - pixiv</a>I use Colab to run it, so it didn't finish and stopped on epoch13.How to use:You can input nothing to generate the character any way, but I do recommend you to use following prompts to generate the outfit:<ul><li>Base: Graf Zeppelin /(Kancolle/), grey eyes, (large breasts:1.3)</li><li>Outfit: black gloves, black pantyhose, sidelocks, peaked cap, pleated skirt, black skirt, necktie, miniskirt, military uniform, iron cross, white capelet, long sleeves, white headwear, military hat, white jacket,</li><li>Try other tags by your own, this has a wide variety of possible clothings, since this model is bit overfitted, don't forget to put those official outfit element into negative prompts, and lower the model weight. I have updated some examples for reference.<ul><li>NSFW compatible.</li><li>I'm not a fan of realistic model, but it seems being capable with ChilloutMix.</li><li>If you find out that it's focusing on chara's feet, you can try to use Controlnet to limit it.</li></ul></li></ul>Settings are what I am using, you can change by your own.<ul><li>DPM++ Series(SDE Karras, 2M Karras, etc.)</li><li>about 20 steps, CFG scale 5~7</li><li>LoRA Weight: maybe above 0.6, try it by your own</li><li>If using highres<ul><li>ESRGAN_4X</li><li>Denoising 0.4-0.6.</li></ul></li></ul>LoRA Train Info<ul><li>based on Animefull-pruned<ul><li>8x low quality 48 images</li><li>16x medium quality 51 images</li><li>32x high quality 33 images</li><li>=&gt; 1 epoch = 1128 images</li></ul></li></ul>

Files with SHA256: 9331d44910db1ac990050b5872b2969c7dbc8956a8d516f67d1d79a6aa22dee4