<h2 id="lora-model-of-the-character-verina-trained-with-simpletunner.-qyv1t72l4">LoRa model of the character Verina trained with SimpleTunner.</h2>v3: Three versions, trained for 6000 steps using ST based on 18 images of 1024x1024 resolution.Essentially resolves the issue of blurriness and can generate characters in non-anime styles.Trigger word: (girl Verina/Verina)Trigger word (anime style): (anime girl Verina)When generating characters in non-anime styles, it is necessary to appropriately reduce the Lora weights to avoid overfitting, which can lead to disordered lines.v3：第三个版本，使用st基于18张1024x1024的图片进行训练了6000步。基本上解决了模糊问题，而且可以生成非动漫风格的人物。触发词：（girl Verina/Verina）触发词（动漫画风）：（anime girl Verina)生成非动漫风格的人物时，需要适当降低Lora权重以避免出现过拟合导致的线条紊乱。------------------------------------------------------------------------------------------------------v2: The v2 version utilizes a new dataset (20 doujinshi images) and a new labeling method (based on the LLaVA 34b v1.6 q4 visual large language model) to train for 2750 steps on SimpleTuner. The learning rate has been increased (to 2e-4) and the scheduler has been modified to achieve better results. Additionally, the LoRA dimension has been reduced to 4 to decrease the file size and minimize the impact on the overall image. Preliminary observations indicate a significant improvement in image quality compared to the first version, and it can effectively reproduce facial ornaments. However, the drawback is that it cannot consistently reproduce hair color. If hair color instability occurs, it is recommended to manually add prompts such as "golden hair" or "yellow hair".v2版使用了新的数据集（20张同人图片），和新的打标方式（基于llava 34b v1.6 q4视觉大语言模型）在SimpleTunner上训练了2750步。增加了学习率（2e-4）并修改了调度器以求达到更好的效果。同时将lora维度降低为4以求减少体积并减轻对画面整体的影响。目前看来，画质对比第一版有较大的提升，同时可以较好地还原脸部发饰。但是缺点是不能稳定还原发色。如果出现发色不稳定，建议手动添加”golden hair“"yellow hair"相关提示词。<hr /><h2 id="v1-model's-performance-is-quite-poor-and-is-not-recommended-for-direct-use.-xd4m7h54w">v1 model's performance is quite poor and is not recommended for direct use.</h2>Verina's Character Lorav1: The dataset consists of 26 multi-angle renders of the official character model in Blender. It was tagged using WD14 Tagger and underwent cleaning. The base model used was FLUX.1 DEV, trained for 3250 steps. The training framework employed was SimpleTunner. <a target="_blank" rel="ugc" href="https://east.moe/archives/1502">Here</a> are the pitfalls I encountered(Chinese).Trigger Word: Anime girl Verina<h2 id="v1-x6wx6tct0">v1模型的效果非常差，不推荐直接使用。</h2>维里奈的角色Lorav1：数据集为26张官方角色模型在Blender里的多角度渲染图。使用WD14 Tagger打标并进行了清洗。基础模型为FLUX.1 DEV，训练了3250步。训练框架为SimpleTunner。<a target="_blank" rel="ugc" href="https://east.moe/archives/1502">这里</a>是我踩过的坑。触发词：Anime girl Verina

Verina_FLUX

v2: The v2 version utilizes a new dataset (20 doujinshi images) and a new labeling method (based on the LLaVA 34b v1.6 q4 visual large language model) to train for 2750 steps on SimpleTuner. The learning rate has been increased (to 2e-4) and the scheduler has been modified to achieve better results. Additionally, the LoRA dimension has been reduced to 4 to decrease the file size and minimize the impact on the overall image. Preliminary observations indicate a significant improvement in image quality compared to the first version, and it can effectively reproduce facial ornaments. However, the drawback is that it cannot consistently reproduce hair color. If hair color instability occurs, it is recommended to manually add prompts such as "golden hair" or "yellow hair".v2版使用了新的数据集（20张同人图片），和新的打标方式（基于llava 34b v1.6 q4视觉大语言模型）在SimpleTunner上训练了2750步。增加了学习率（2e-4）并修改了调度器以求达到更好的效果。同时将lora维度降低为4以求减少体积并减轻对画面整体的影响。目前看来，画质对比第一版有较大的提升，同时可以较好地还原脸部发饰。但是缺点是不能稳定还原发色。如果出现发色不稳定，建议手动添加”golden hair“"yellow hair"相关提示词。

Files with SHA256: 750d60845a64ea26be347a1d657f38fb6086c3f23961cdce0054de0382c86a94