The motion in the lora is good but there's alot of texture smearing and it gives every character really long eyelashes. I made a patched version that mitigates this scaled attn1.to_v in transformer blocks 0–15 down to 20% and attn2.to_v down to 15%, which prob handles the appearance from the image conditioning in I2V. Also scaled attn2.to_q and to_k down to 30%. Eyelashes mostly gone, motion and 3D style mostly intact. The washout/color fade on high-motion objects is still there though — tried hitting mid and deep block to_v and to_out projections but that just broke the lora, so it's probably too entangled with the motion training to fix without retraining. Happy to share the patched lora if useful.

After a duration of more than 10 seconds, there will be delayed ghosting when the head moves up and down, and I feel that the dynamic effect of ltx2.3 itself is not very good

possible to use it with best faceswap lora on humans ?

I'm just starting out with ComfyUI so please be kind! Can someone provide me with a workflow for image to video that loads action loras like this? Big help! Thank you!