Turbo LoRA distillation of Kirazuri (Ernie)

Version 0.1

Trained with On-Policy Distillation in DiffSynth Studio for 5000 steps, ~13.90s/it

Recommended usage:

25 Steps

CFG 2

Training target was 8 steps, but this preview version is only partially baked.

The recommended settings reflect this, and the generations are fairly stable at half the base's 50 steps and CFG of 4.

This will need a larger LoRA rank and longer (at least 3x) training, which I plan to try once the Kirazuri (Ernie) base itself has undergone more training.