通过添加正则化与修改训练参数让lora更优秀,与前两个版本相比这个版本无需1.5权重即可还原角色(正常权重即可还原角色),拥有更好的提示词响应和还原。请注意,训练分辨率较高,最佳分辨率为1920x1080,最低为1344x768,推荐分辨率比例为16:9,但其他分辨率比例在测试中依旧有着不错的表现。
By adding regularisation and modifying the training parameters, LoRa is even better, and compared to the previous two versions, this version can restore the role without 1.5 weights (normal weights can restore the role), and has better prompt response and reversion. Note that the training resolution is high, with an optimal resolution of 1920x1080 and a minimum of 1344x768, with a recommended resolution ratio of 16:9, but other resolution ratios still performed well in the tests.