Latest update is the version division between the old artist works and the newer ones.
Pixiv link to original style artist: https://www.pixiv.net/en/users/49675420
Description
This version should output slighly better results as it was trained with 768 resolution and looks a bit less overfried.
The recomendation is to use sd-84k vae, euler a sampler, animesharp hires upscaler with different denoising strength depending on the camera position and halving the text encoder strength(TEnc weight) with additional networks extension(set it to 0.5).
If the network stops following the prompt consider the recomendation to disable lora text encoder at all(set it to 0). Its also recommended to do for combinations with another loras.
FAQ
Comments (2)
There's nothing special about bf16 precision, it is a lower decimal precision than fp16 and deep learning models' values are meant to stay within the range of [-1, 1], so the additional exponent space is useless :D . It works fine even on Mac M1
I've tried to train with bf16, fp32 and fp16. Only the fp16 version of loras with high alphas has the tendention to output 0 valued tensors, even the lower alphas making some numbers of them, while with bf16 and fp32 there is no zeroed tensors at all even with high alpha value.