Female Warrior from Isekai Yarisaa Update v1b (4th June 2023):<ul><li>Experimented with Differential Learning to remove the anime art style. (<a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts/pull/542">https://github.com/kohya-ss/sd-scripts/pull/542</a>)<ul><li>First I had to train a style model from the anime screencaps without the character I wanted to train. The problem was that I hadn't trained a style model before. If I remember correctly, there are mainly two ways to train one: either training U-Net only or auto-tagging the dataset with a lower threshold without any pruning. I decided to try the first option. In the end, I am uncertain if the style model has replicated the style well enough or if the character model is style-neutral enough, but the character model definitely looks less animeish.</li></ul></li><li>Added two training parameters for experiment: scale_weight_norms and network_dropout.</li></ul> Prompts:-<ul><li>Trigger Words: femalewarrior_olis, red hair, long hair, green eyes, huge breasts</li><li>Outfit: earrings, choker, o-ring, shoulder armor, bikini armor, underboob, armlet, corset, gloves, black panties<ul><li>Extra: over-kneehighs, boots</li></ul></li><li>Negative Prompts: (censored, mosaic censoring, bar censor, convenient censoring, pointless censoring:1.x) (please experiment with the weight. Sometimes it's better to simply change the seed...)</li><li>Extra: (please refer to my sample images) (I ACCIDENTIALLY changed the Clip Skip in some of the generations for some reason. Please stick with 2.)</li></ul>Weight: 1 (you may experiment yourself)All my models were trained with 512x512 pixels. For optimal results, please upscale your generated outputs with your preferred methods. My ninth model.This was probably the only time I took a request. Anyway, I continued to experiment with finetuning. The results weren't bad, I guess.Hopefully some of you will like it.

Female Warrior 女戦士 | Isekai Yarisaa

<ul><li>Experimented with Differential Learning to remove the anime art style</li><li>Added two training parameters: scale_weight_norms and network_dropout</li></ul>

Files with SHA256: f8f45e117705973f51537651ca5823b287109c8efc6159f580cb8b11a7b9ab94