This is just the base kijai workflow released on 8/7 with the new lightning loras also released 8/7, with the addition of GIMM interpolation. RIFE47/49 is also there, if you want to use it just hook it up and enable it. The first time you use GIMM it'll need to download, but it should be taken care of with the workflow. Note this uses sage and torch compile, which speed up generations by a lot, but you can disable them if you don't want to use them. I'm assuming you put your models in WAN22, if not you'll need to update that.
Description
FAQ
Comments (19)
I assume these are the new lightning?
yep, that's it
The new lightning models turn everything slow mo, a real shame.
yeah that is true. i find that less so if i use 720 or 1280 resolutions, but yeah i've on and off had that issue since wan 2.1 without any loras. i think with how capable the 2.2 models are its worth trying things again with no loras (other than lightning)
cant you just change the length/frame rate math or interpolation multi?
jwcoop3r631 sure, but there's no way of knowing in advance if it'll be a slow generation or not. better to do it in post with with capcut (or something better), to speed it up.
How to disable sage and torch compile?
its really worth installing both, they're fantastic. that said, to disable torch compile you can just bypass the #35 node (WanVideo Torch Compile Settings) and in both of the Wan Video Model Loader nodes (#22/#72) select "SDPA" instead of sageattn. But yeah I'd recommend getting Triton/Sage Attention working instead, its much faster.
I get this error, but I can't see any obvious fix.
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 68, 21, 45, 45] to have 36 channels, but got 68 channels instead
https://www.reddit.com/r/StableDiffusion/comments/1mc2gkq/given_groups1_weight_of_size_5120_36_1_2_2/ try this thread? or try it without sage attn and torch compile.
@IrakTheOracle Thank you! I actually found that one and it seemed like the VAE had to be Wan2.1. Now it gets stuck at the sampler stage and I didn't see any sign of progress for 30 minutes. A good old reboot might do the trick :P
May I ask what GPU you use?
@PDMS i use a 5090, additionally i have 192 gb of ram, so i might be a poor test case.
@PDMS if you get stuck and its actually generating though, try adjusting the block swap to be larger (30-35). additionally lower the resolution. the biggest i ever do is 1280x720, even at 5090. finally you can lower the frames to 41 for prototyping, but i try to avoid that as it can mess with the stability too much, but should be an easy way to confirm if it'll generate at all.
@IrakTheOracle Thank you so much! I will try your suggestions. Just purchasing a 5090 feels inevitable though xD
I run RTX 4060 Ti with 16GB VRAM and 48GB RAM besides that.
@IrakTheOracle Block swap did the trick! Thank you!
We cant add loras ?
only 5 seconds?
Got it working on my 3060, getting torch, sage, and triton was a pain for all them to play nice. Not the workflow fault though. This is super fast to get a video.
This workflow is neat. I get excellent results in 720x1280 - 8 seconds O_o. The details are insane and the action is ultra smooth with NO degradation. OMG a must have !