<h2 id="hoshino-ichika-f.1-dev-lora-model">Hoshino Ichika F.1 Dev LoRA Model</h2><p>A LoRA Model based on the base model <strong>FLUX.1 Dev.</strong></p><p>I use two different kinds of images for training: Cards (not transparent) and Cutouts (transparent). <s>So the background will be single-color, not flexible, I think. </s>So you can control your background freely, I think.</p><h2 id="the-usage">The Usage</h2><p>Well, The LoRA can be triggered by the single word 'hoshino ichika', However, I still recommend u to add 'hoshino ichika, long hair, blue eyes <strong>(or grey eyes)</strong>, black hair <strong>(maybe 'blue hair' is also ok?)</strong>' first to your prompt to make sure the characteristics are correct. (If u use <strong>ComfyUI-nunchaku</strong> workflows, u even do not need the so-called trigger word) And it's recommend to use tags in the Danbooru Dataset only.</p><p><s>Also, I only tested the case that the weight of the LoRA is 1.25, and I don't really know if lower value can still make the output normal. (You can test it yourself and tell me in the comment area)</s></p><p>In my tests with the help of <a target="_blank" rel="ugc" href="https://github.com/nunchaku-tech/nunchaku">nunchaku</a>,</p><p>I can use the LoRA with the weight of 1. And don't need to add the LoRA file name in the prompt input.</p><h2 id="the-introduction-of-the-character">The Introduction of the Character</h2><p><a target="_blank" rel="ugc" href="https://www.sekaipedia.org/wiki/Hoshino_Ichika">Sekaipedia</a></p><hr /><h2>一些基本的训练参数</h2><p>素材以人物卡面为主（Cards+Cutouts），其中Cutouts是透明的但Cards不是透明的（</p><p>训练脚本是直接采用sd-scripts的sd3分支，并受到fluxgym写script的启发，自行手搓脚本和数据设置而成。</p><p>考虑到本人笔电配置不佳，所以每次训练的效率不高，并且迄今为止找不到办法有效降低avr_loss的办法 <s>(除了拉高steps，但那样更费时间)，</s></p><h2 id="lora">怎么用这个LoRA</h2><p>跟正常使用SD 1.5的LoRA一样，用FLUX.1 Dev的底模，然后使用触发词'hoshino ichika'触发LoRA（小提示：如果你采用了跟我一样的ComfyUI-nunchaku工作流去加载这个LoRA的话，甚至连这个所谓的触发词都不用写，直接开始写该角色的主要特征），权重默认为1即可。当然，为了保证角色的基本特征被正确触发（<strong>我通过删减一些tags来固定基本的人物特征，实际证明可能没什么太大用处（？）</strong>），推荐加上"hoshino ichika, long hair, blue eyes (或者grey eyes), black hair (可能blue hair什么的也行？)"这些特征词，并且尽可能只使用DeepBooru词库的词语。</p><h2>面临的挑战</h2><p>这个LoRA由于只是对这个新角色的训练，因此在肢体等一些细节上仍有很大的不足。遗憾的是，f.1并不像sd 1.5那样可以随意加入embeddings来暂时克服肢体问题（可以考虑加入ControlNet来控制图像走向，可能有效果）。我的评价是，反复抽卡吧(</p><p>（当然如果你有很好的解决方案可以在评论区指点一下，本人感激不尽x）</p>

Files with SHA256: e651283ea7473304e2d49f965731f6328e460fd3f837004b548d5f0804253949