Anima acc lora
A 16-step distillation LoRA.
Settings: CFG 1, 16 steps.
Works on any base ckpt.
It's not a better "turbo" LoRA, it's just another one, with different approach.
EDIT: About cover images:
Cover images are from my RDBT model (private version), highly stylized drawing, to demonstrate how distillation affects styles and details (losing details, there is no free lunch. And yes, I deliberately demonstrate how bad distillation could be)
When is distilled model recommended:
If you are looking for production level drawing and textures, or have a beefy GPU. Don't use distilled model.
If you are here for general anime styles, then this 16-step acc LoRA is OK. Most of the time you can get identical output, but 4x faster.
Difference between official turbo model:
More entropy. More details. More natural.
No detail/quality amplifier. Doesn't overwrite styles with a unified AI slop style. (TBH, I think this might be a bug in the official turbo. It adds artifacts and patterns. Overcooked probably.)
For users from my RBDT model:
I will stop releasing distilled RDBT ckpt from now on. I previously thought that if I distilled directly onto the fine-tuned ckpt, I could retain more detail. In reality, the difference is minimal to none. Therefore, I'm not wasting computational resources on each version.
v2.4 is trained/distilled with more compute and a much larger dataset. It has much better details than previous v2.3 distilled ckpt.
Description
FAQ
Comments (7)
I think 16 steps is too much for official turbo lora. This explains burnt look.
Yeah I think the official Anima Turbo model and Lora only needs 8 - 12 steps.
number of steps doesn't matter.
it's a diffusion model. eventually the model will converge to a fixed point. More steps == more accurate
if you think lower steps is somehow better, means the converged point is so AI-ish, that it would be better just not let it fully converge. Means the turbo LoRA itself is fked up.
@reakaakasky I don't think that's his point. He is saying that your comparisons images are not conclusive, because with 16 steps you make the official turbo lora look ridiculous. If you want to do 16 steps with it, then don't apply it at full weight, but like 0.8 or whatever.
@deitychaser yeah, I got the point. I previously also used the official turbo a lot. But I think it's training target is a little bit off. E.g strength> 0.3, model starts to ignore brightness prompt. >0.8, there are artifacts and patterns. I was trying to demonstrate this.
strength 1 + enough steps = model training target.
Thanks for the continued work on Anima!
Do you have plans to release non-distilled/"base" versions of Acc? I prefer the non-distilled versions of RDBT, so I hope you'll continue to release them for Acc as well.
there is no "base" version of acc LoRA. it's just a standalone distillation LoRA. There is no fine-tuning. Cover images are from rdbt, the quality might be a little bit misleading.
From now on I will only release the base version of RDBT.