中文：第一次炼XL的lora，这效果是真的逆天，学习率拉到以前1.5想都不敢想的高度都没有过拟合，炼丹，爽！强度为1即可注意：1、使用了FP8模式训练，使用FP16跑图的话可能会出现未知问题。2、由于数据集过少和重复度过高，大部分情况下需要添加特殊的（不可描述）触发词才能获得对应的画风。3、由于XL模型的特殊性，可能要多次尝试不同提示词才能获得想要的结果。4、测试下来是基本能够还原画风的，但由于某些原因我不能直接发布在C站上。5、目前只测试了Pony原模型，其他衍生模型可能会出现崩图。English:First time to train lora on SDXL and was astonished by its performance, lora still havnt been overfit as i set the lr at a high rate which couldnt be imagined for sd1.5. I can do this all day! XDweight 1note:<ol><li>Use FP8 training mode, there may be unknown problems of FP16 .</li><li>Because of the SMALL SIZE of datasets and the high repetition, in most cases you need to add special (forbidden in Civitai) trigger words to get the corresponding style.</li><li>Because of the particularity of XL model, you may have to try different prompts several times to get the results.</li><li>the test down is basically able to restore the painting style, but for some reason I can not published on the Civitai.</li><li>Only the original Pony model has been tested so far, and other derivative models may collapse the picture.</li></ol>

Files with SHA256: 069f3347a4472bc6375a0968ae9c3bef662f5457189527ff3c92a962d599808b