Workflow is not mine. It's here if anyone wants it
https://civitai.com/models/732926/flux-lora-block-weight-tester>>85283128>Faster training and less impact on the overall model?Yes. Not training double blocks also saves a lot of memory
>>85283562Try changing "Noising Bias" to 0.7 or 1.0
it should look similar to shift=3 in kohya scripts
https://github.com/kohya-ss/sd-scripts/tree/sd3?tab=readme-ov-file#distribution-of-timestepsDefaults in onetrainer are equivalent to shift=1. It won't learn details.
Shift 1, 2, 3 looks like this from top to bottom:
https://files.catbox.moe/soxxxz.jpg