Thanks for this update! with this lora you can get good results even with 10 steps + cfg=1.0.

v0.12 = insane quality improvement. Both anime and realistic. Might be a game changer.

Sorry to hear that you won't update anima. Not a professional. But it sounds strange that they can also sell other people's free models.

Always looking forward to your next model!

Thanks to reakaakasky for pointing out the licensing restrictions of Anima.

Thanks for this lora, it is working extremely well in my tests! It is a shame what the license terms specify so I completely get why you might not continue with training this lora (I would still appreciate if you do!).

I think the license along with some questionable choices regarding the model architecture could prevent anima from becoming a true illustrious successor, it's a shame.

Well, finding out about this model's capabilities and then reading that description sure did a number on the hormones.

I wanted to say thanks for the wonderful LoRa. The final result is much better and I almost never turn it off. The only downside to using it - some @named_styles are weeker.

For those who don't know yet, you can speed up the generation process even more by adding "CacheDiT Wan Accelerator" node in ComfyUI (between KSampler and LoRa). For me, using a rdbt-LoRa or CacheDiT, the speed is about 30 seconds. Using both - 15 sec. (None - 50 sec).

When using this node the final result may differ, but usually the similar result is obtained with a small changes in the prompt or steps in euler_ancestral.

Settings that suited me: warmup_steps=3, skip_interval=2.

Important: Manual installation is required, but instructions are available on gihub: Jasonzzt / ComfyUI-CacheDiT