2023/3/28 update
Use Min-SNR Weighting Strategy to control loss and accelerate convergence.
Introduction
A conventional LoRA model trained with ~90 images of Olga Discordia from Kuroinu. 0.6 ~ 0.8 weights are good. The preview images are generated using 1.6 version (original) YesMix. Now she can play with Celestine Lucullus.
Cautions (Must Read)
Some preview images are generated using hires fix. Thus, if you wanna generate the preview images, pls download the original preview images and analyze them using 'PNG INFO' in webui to get all generating information.
The maximum of generating resolution should not exceed 1024 theoretically, otherwise extra/deformed head/limbs will be waiting for you :). Try to use 'hires fix' to prevent this problem and help to generate higher-resolution images.
For 784mb VAEs (NovelAI, Anythingv3, Orangemix, Counterfeit) users, make sure add '--no-half-vae' in command line of webui to prevent generating black images. See voldy's log for the details of edit method. For high-performance GPU, make sure 'medvram' and 'lowvram' mode are disabled.
Description
Use Min-SNR Weighting Strategy to control loss and accelerate convergence.
FAQ
Comments (4)
How can i use Min-SNR Weighting Strategy
is there an extension?
No, it's a method that can accelerate convergence during training, not an extension.
FYI https://github.com/kohya-ss/sd-scripts/pull/308 has the feature.
Why can't the uploaded images recognize this Lora?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.