I trained on a dataset of 300 images of not very high quality, mostly from JAV. The chilloutmix checkpoint was used for training. The optimal weight is 0.8, but it depends on the checkpoint. For some reason, it is much harder to get a decent result on some checkpoints, curiousmerge2.5D is the easiest one I tested, while chilloutmix, which was used for training, is more difficult. With high weight, have some problems with unexpected censorship and icons in the corners of images