EDIT: (currently using the 1ksampler workflow with this new version of the acceleration lora: https://civarchive.com/models/1585622?modelVersionId=2014449)
2in1 - Includes an updated version of the 1ksampler workflow and a new 2ksampler workflow.
fun-reward mps lora
Self forcing accelerator lora:
https://civarchive.com/models/1713337?modelVersionId=1938875
still has my custom comfyui themes in the zip file.
DISCLAIMER! You need the fun-reward-mps and the self forcing acceleration lora for these to work properly.
Update: with the 2ksampler workflow i managed to squeeze it to a total of 4 steps:
1st Ksampler - total steps 4 - steps 0 to 1 - cfg 3.5 - return with leftover noise DISABLED
2nd Ksampler - total steps 4 - steps 1 to 4 - cfg 1 - add noise ENABLED
V4.1 - Removed florence, it's kinda useless with this setup. Moved shift before sage attention, i think it's supposed to be like that, atleast it feels better now. Still contains my custom comfyui themes as an extra.
Wan gguf workflow using one of the self forcing loras to use low step counts.
Use whatever gguf model you want, i'm using the Q5_K_M.
Self forcing accelerator lora:
https://civarchive.com/models/1713337?modelVersionId=1938875 (i'm using this one)
https://civarchive.com/models/1585622?modelVersionId=1909719
Also includes my custom themes for comfyui, to install it click the kogwheel (settings) - Appearance - color palette - import: (Astraali_RedLINE.json / Astraali_DarkLINE_1.json / Astraali_BlueGOO_4.json)
Description
FAQ
Comments (3)
I've been messing about with your workflow and its very nice! Results are great.
I have a question though, which may be something silly I'm doing. But everything is very slow motion.
I've prompted against it and set shift to like 8-9.
If I set it to 16/32fps and double for the upscale to 32/64fps I get normal movement, but the clip will go from 7sec to like 3.5sec.
Is there a way to get the faster movement in a 7sec video clip. Or is this just how it is?
Yea i've noticed the same thing with the movement being slow, and i think it's just what the self forcing lora and low cfg does. I think doing 4 steps instead of 5 increases the movement, you could also do lower shift, like 3-5. Also because of the lack of cfg, you can try put more weight on the prompts to compensate. Some loras help a lot with speeding up the movement.
Also i don't know how much sampler/scheduler affect motion, but maybe try uni_pc + simple instead of LCM + Beta