In a few days of debugging, I basically found the training parameters and patterns of sd3.5large. This is a sample.

The current progress can basically achieve 12GB graphics card memory for training SD3.5 lora

It is necessary to freeze some layers, but in reality, freezing can achieve better results

Under Lora's small-scale image count, training all layers with SD3.5 may cause some issues.

But at relatively large image scales, all layers can be trained.

The one click package of training scripts has been published on my personal media: https://t.zsxq.com/PIme8

Next, we will update the frozen strategy and various training plans in personal media.

The training script comes from: https://github.com/kohya-ss/sd-scripts

Thank you: hhy yizhou