Why are there k_img and v_img in the low noise lora?
Wan 2.2 does not have these modules. Do you also plan to use this lora for Wan 2.1 I2V?
This model focuses on training for high noise. For low noise, using the LoRA from Wan2.1 directly still yields good results.
What are the recommended step counts? Up to 10 steps high noise tested and it has poor prompt adherence and subject ghosts/teleports.
Seems OK with two steps of high without lora just not sure that's the goal?
ghosts/teleports too. only replace the original two lightning loras
What are the recommended step counts? Up to 10 steps high noise tested and it has poor prompt adherence and subject ghosts/teleports.
Seems OK with two steps of high without lora just not sure that's the goal?
I think these two models are similar to lightx2v/Wan2.1-I2V-14B-720P/480P-StepDistill-CfgDistill-Lightx2v, as they all use 4 inference steps.