This is a revised version. I trained the model again, updated the training dataset, and found that the results are slightly better than the previous version. It can more accurately preserve facial expressions such as winks, reduce the number of generations needed to get a good result, and reproduce the original image better than the previous model even without prompt optimization. However, I consider this model to be a minor update, so I named it v2.1. Feel free to test it yourself and use whichever version you prefer.The released model was trained on an RTX 5090 for 9,600 steps. It still has not been specifically trained on NSFW content. If anyone has good NSFW training images, feel free to discuss them with me (After achieving full expression preservation, I may consider working on the preservation of various liquids, physical compression/deformation effects, and the realistic material reconstruction of real-world objects.这是一个修正版本，我又进行了一次训练，更新了一下训练集，然后发现训练结果略微比上一次好，在一些wink等的表情上会更准确还原人物的表情，减少抽卡次数，且在没有进行提示词优化的情况下能比上个模型更好还原原图。不过这个模型我认为应该作为一个小幅更新，所以叫做v2.1。你们可以自己测试一下然后凭你们自己喜好使用。 发布的这个模型是用rtx5090训练9600步得到的。依然没有进行nsfw的针对训练。如果有人有比较好的nsfw训练图可以和我讨论( 后面做到表情的完全还原之后考虑进行各种液体的还原以及物理挤压以及各种真是物体的材质还原等等。

Files with SHA256: 6305294dc00b2b61aca0edb5ddd0bb354a72f98e37194fe8a64a5deab48845ca