Among other improvements, the new defaults set --flux_guidance_value=1
, removing the need to use CFG nodes at inference, reducing generation time and improving image quality of LoRAs slightly.
Changelog: https://github.com/bghira/SimpleTuner/releases/tag/v0.9.8.1
Sample LoRA: https://huggingface.co/ptx0/flux-dreambooth-lora-r16-dev-cfg1/blob/main/pytorch_lora_weights.safetensors
I still havent figured out nice lora training settings for like 108 images.