Turbo LoRA distillation of Kirazuri (Anima) v4.0
Important: This LoRA is not expected to work with other base models.
Generation Settings:
4 or 8 steps (see model version names), CFG 1
sampler/scheduler:
res_2s + bong_tangent
res_multistep + simple
dpmpp_2m + sgm_uniform
Resolution:
1536^2 recommended for best results
Otherwise, please refer to recommendations on the Kirazuri (Anima) base model card.
Training details (Latest - v4.0):
Created with DP-DMD diversity-preserved few-step distillation.
Training dataset of 959 images generated by the teacher model Kirazuri (Anima) v4.0 with fully synthetic and partially synthetic images generated with img2img.
Generated at 1536^2 with a diverse set of prompts intended to increase the style coverage and diversity of the base model.
Selected only generations without faults for training (i.e. anatomical or compositional errors).
Thanks to sorryhyun for the trainer: https://github.com/sorryhyun/anima_lora
Description
FAQ
Comments (2)
i feel like 4step might be a bit too low, it feels like the images got VERY generic compared to just using kirazuri+12step turbo.
Thanks, I am still learning a bit about this.
If this distill feels a bit different/generic it might be due to having used a tiny subset of the real data used to train the base model as the dataset for distillation.
This means it also seems to fixate on differences in the real data as a mathematically impossible target, so any artifacts or minute details like background noise receive extra attention.
From what I gather a synthetic dataset of the teacher models outputs is needed for a distill to accurately produce similar outputs in fewer steps.
I might give this a try, though I wasn't really planning to go down a distillation rabbit-hole. (ᵕ—ᴗ—)