<h1 id="turbo-lora-distillation-of-kirazuri-(ernie)">Turbo LoRA distillation of <a target="_blank" rel="ugc" href="https://civarchive.com/models/2994871/kirazuri-ernie">Kirazuri (Ernie)</a></h1><h3 id="version-0.1">Version 0.1</h3>Trained with On-Policy Distillation in <a target="_blank" rel="ugc" href="https://github.com/modelscope/DiffSynth-Studio/blob/main/docs/en/Training/Distillation.md">DiffSynth Studio</a> for 5000 steps, ~13.90s/itRecommended usage:25 StepsCFG 2Training 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.

Files with SHA256: 9cbc98b757f36986e68362ab1d8353966aee94bcd2e2bd2c614bcf148213cb31