Innovate Futures @ Benji
patreon
Wan 2.2 Animate (Long Length Gen Looping (Advanced Pose) Ver.20250924) - Additional For Patreon Supporters
🕑 Added 2025-09-24 15:09:57 +0000 UTCIn the previous public post, I have post the general Long Length Gen Looping method for Wan 2.2 Animate native node. https://www.patreon.com/posts/139637572
In this post, I am going to post additional enhancement for this workflow.
1 - DwPose Extract
If you have low VRam , or you are processing very long length ,e.g: over 2 minutes duration.
I suggest to have another workflow that only take care of extracting DWPose and Face video first.
It helps a lot to reduce runtime memory when using animate workflow.
Because having the DwPose proprocessing like this , it store all pose frames in temp.
Instead I save pose and face video with another simple workflow in step 1.
Then directly select the video , pose video, and face video in this group.
One more thing it does is that, it prevent ComfyUI sudden pipeline broken or connection lose.
Cause I noticed new ComfyUI version do cause this issues.
So, work efficiency, you don't have to start all over again... if ComfyUI broken in the meantime you generate long length video.
You might have to...
loading video frame....
extracting DW poses..... (and my video are all over 1000 frames...)
Detecting face.....
So check it out and have fun :
Comments
Benjamin
that's very limited intel., it's very hard to determine what's wrong with your ComfyUI.
Pomoe
I get a completely black video output. What could be the cause?
Benjamin
For quality wise, if a 3090 can go for Q6 or above. the image will looks better? About OOM, you have reach the max of it for that hardware config. I heard one of the 3090 users in our Discord mentioned similar experience as well.
Brandon Waters
Ok so played around with the split workflow. I was able to get dwpose to run if i lowered the resolution to 640 x 640 and the image resize from 512 to 256. The other workflow, i can get to about 860 frames before it goes OOM. I have a 3090 with 24GB of Vram, using Q5 GGUF's for the clip and the diffusion model. Any ideas?
