<h2 id="heading-106">Trained on animagineXLV3_v30, generate images of All Out Attacks execution scene of Persona 5.</h2>Recommended weight: 0.8 is recommended.(Although a higher weight generates a style more similar to the original, it tends to introduce noticeable errors. Using 0.8 provides a more balanced approach.If you want to restore the style, use 1.The first picture on the cover is 1, and the others are 0.8)The dataset is trained at a resolution of 1152*1152, comprising 53 images.The optimal generation resolution is 1536*864, which is the aspect ratio of almost all datasets.Trigger words: All Out Attacks, silhouette, English text, mask, glove.(the primary trigger word is All Out Attacks; this is my first time training the Lora style, and there is no comparative data in the dataset. During testing, it was found that adding or not adding these trigger words made little difference. Other trigger words can be added as needed.)The labeling method involves full tagging + trigger words.Slightly adding some character tags can help alleviate quality loss, such as red eyes, red hair.After the training was completed, I learned that animagineXLV3_V3.1 had already been released...===========================================================<h2 id="heading-123">基于animagineXLV3_v30训练，生成女神异闻录5的总攻击处刑画面图像。</h2><h2 id="heading-124"></h2>推荐权重：推荐0.8。(虽然更高的权重生成的画风更像，但是会出现明显错误，使用0.8来说更均衡。如果更想要还原画风那就使用1吧。封面第一张图是1，其他是0.8。)数据集以1152*1152分辨率进行训练，有53张。最佳的生成分辨率是1536*864，几乎所有数据集都是这个比例。触发词：All Out Attacks,silhouette,english text,mask,glove (主要触发词是All Out Attacks，这是我第一次训练画风lora，数据集没有对比数据，测试时发现加不加区别不大。另外几个可以根据需要填写。)标注方式是全标+触发词。稍微加一点形象tag可以缓解质量损失，如red eyes,red hair。在训练完成后，我才得知animagineXLV3_V3.1已经发布了...

Files with SHA256: fcc1304aa9fa8158cc8e1203662b0d32e471379f1f43ff1eeb719ed125638bfe