Waifu Painting (JWW)版本v4 LoHa Epoch 10 (ID: 34451)

Only for SD2.x models such as Waifu Diffusion 1.5, will not work SD1.x models such as Anything, AOM, Counterfeit, and etc.

A painting style LoRA trained on Waifu Diffusion 1.5 beta2, with training data from images of paintings of John William Waterhouse (69 files gathered from wikipedia) with almost no tagging.

Trained with noise offset = 0.1, so it will conflict with other LoRA without noise offset.

Also, it doesn't generate pure landscape well, as John William Waterhouse is purely a "waifu" and mythology painter, all of his paintings contain human subjects (I've tried to fix this problem with v2).

Training parameters:
For v1: Learning rate is 1e-5 for both unet and text_encoder, scheduler constant with warmup. Image repeat = 10.

For v2: Same learning rate, but with augmented data. I cut out the parts of paintings that only have background (for example: https://files.catbox.moe/ad6tyo.png) and add them to the training sets.

For v3: Instead of cutting background part to new images, I erased the subject instead. But it doesn't seem to work well.

描述:

Recent webui and extensions updates changed a lot of things, so I decided to retrain a new version. This time I remove the word "painting" from trigger words, because sometimes it mistakes it as actual painting (object) instead of painting as medium. I also reduce tagging as simple as possible (no tag is better than confusing tag).

Also now the extension https://github.com/KohakuBlueleaf/a1111-sd-webui-locon is working properly, so ridiculous high weight (such as 9 or 10) isn't necessary.

训练词语: by john william waterhouse