>>936008569
>>936008723
>>936008743

yeah! its very setting dependent.
self-forcing is very efficient and fast
unet ggufs are faster but have a bit lower quality
causvid lora helps
teacache and sageattention (no fucking clue what these are technically but they make it go fasterer)

im running the full 14b WAN 2.1 bf16 model. I generate 81 frames at 480 at 32 steps. I am using a 4090 with 24gb vram. actual utilization comes out to somewhere around 18ish
the 5 second (81 frame) video takes around 420-560 seconds per gen

i havent tried self-forcing yet but i shall.
right now i use sageattention and causvid to bring the render time down
im happy with the results im getting with 32 steps and i have the hardware so i might as well. but im sure ill swap it up at some point, especially because im eyeing VACE