See also lightx2v / Kijai / Wan2.1, Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1
Note:
•Monday, August 11, 2025 at 3:53:02 PM UTC
Honestly I don't really know, it feels like they used different method to train it and it's just not as good, it doesn't feel like the self-forcing LoRA at all, worst part of this one for me is that it has a clear style bias, it makes everything overly bright, you can't really make dark scenes at all with it, and it tends to look too satured.
I'm mostly still using the old lightx2v by scheduling LoRA strengths and CFG. The new LoRA can be mixed in with lower weights too for some benefit.
There seems to be an official "Flash" model coming from the Wan team as they just teased it, hoping that will be better.
>
--
Fabulous-Snow4366:
testing it right now using (fp8, 8 steps 4high 4 low, 121 frames, sage attention on), on my 5060ti, its roughly twice as fast as without the Loras and sage attention, around 30secs/it compared to 75secs/it. BUT its still slow-motion galore, reducing movement by a lot.
Any_Fee5299:
121 frames is for 5B, this LoRA is for A14B version. Use lower (0.5-0.95) str on high
>