<h2 id="an-anime-character-f.1-d-lora-model-nb0mnessk">An Anime-Character F.1 D LoRA Model</h2><p>The character is Ayase Eli who comes from the series called "<strong>Love Live! School Idol Project"</strong>.</p><h2 id="the-solution-for-lora-training-at7jxeqi4">The solution for LoRA training</h2><p>Well, I just use the <strong>sd3 branch</strong> of the repo <a rel="ugc" href="https://github.com/kohya-ss/sd-scripts/tree/sd3">sd-scripts</a> to train it and I tried to write the dataset config file and the parameters of training unet scripts MANUALLY. However, I couldn't find a way to obvoiusly reduce the average loss and due to the <strong>BAD performance</strong> of my laptop, I may could not upload other characters LoRA during a short time.</p><h2 id="how-to-generate-images-like-the-preview-images-eduq9tx8j">How to generate images like the preview images</h2><p>OK, fine, I am not good at choosing the best generating parameters. I just tried the Weight of LoRA 1.25, and I'm not sure what the lowest weight value can generate a good image enough. You can strictly follow my workflow nodes and the prompts showed in the preview images.</p><p>I recommend using the words in the <a target="_blank" rel="ugc" href="danbooru.donmai.us">DeepBooru</a> dictionary. If u want the Florence 2 version, please tell me.</p><p></p><hr /><p>这是我个人第一次尝试上传自己训练好的F.1 D LoRA，训练参数上受到fluxgym写的scripts的启发，决定自己手动调用sd-scripts的sd3分支来完成LoRA训练。</p><p>至于使用的关键词，我推荐使用来自DeepBooru的词汇，因为目前暂时没有试着用Florence 2辅助打标 <s>（太麻烦）</s>。</p><p>触发词很简单，就是一个ayase eli就完事，但是为了保证角色特征一致性，建议还是稍微写一下眼睛和头发的基本特征（也许也可以不用？）：</p><pre><code>&lt;lora:ayase-eli:1.25&gt;,ayase eli,1girl, solo, blue eyes, long hair, blonde hair,</code></pre><p>考虑到本人的笔电性能较差，所以可能不会短时间内大批量训练很多的角色LoRA，见谅！</p><p>至于本人使用的workflow，我这里为了生成速度和更低的硬件资源占用，花了大约一整个晚上的时间，安装好了<a rel="ugc" href="https://github.com/mit-han-lab/nunchaku">svdquant</a>这个很强大的工具，如果你想用stable-diffusion-webui之类的，或者用ComfyUI的传统workflow，只要能兼容使用FLUX.1 Dev这个Base Model，就可以用这个LoRA！（确信</p><p></p>

Files with SHA256: 727a648f9ecea14adef2f68f8b8f5431eb865625e6fa5d85e3f375d7d518898f