following my simplest WAN 2.2 workflow with loras (turbo and NSFW), i did very similar workflow for double length videos. It also works perfectly with 16gb VRAM.
just as previous workflow, this should work on any comfyUI installation.
You will need to download all models, vae, clip and loras yourself unfortunately.
how/why set it up.
first part is set as a normal i2v clip. You want MAXIMUM 5seconds length if you want to change the details between first and last frames. if you go longer, the scene WILL start going back to frame 1.
for the second part it will use last frame as it's first. again, 5s if you want to avoid going back, depends on scene/prompt. 7sec in default settings.
it requires a little luck to get last frame without high movement, or the quality of second clip will drop.
use 4steps for low changes in the background and 6 if quality is a bit low.
have fun.
turbo (lightning) loras are here: https://huggingface.co/lightx2v/Wan2.2-Distill-Loras/tree/main
version might be different when You browse it.
Description
FAQ
Comments (6)
I've found using last frame to be frustrating as most gens tend to introduce brightness/color shifts that make the seam poor. I've tried some color matching, but haven't found a solution that looks great yet.
I noticed it too, but only sometimes. Besides I'm not ambitious enough to actually do multi clip videos ;p
@logoth I've been working on a multi-clip workflow, and haven't found a solution for this part yet. I'm mostly relying on edit LoRA's which helps, but for longer shots I've not cracked it. (The video I made from your image uses it: https://civitai.com/images/124593955) You can see the color shift at the end of the first clip. It's quite annoying.
You're looking for SVI 2.0 Pro. If used properly, you can do 20+ seconds without visible color shift. https://github.com/vita-epfl/Stable-Video-Infinity
There are workflows and examples all over this site utilizing it.
I don't recommend SVI Pro, myself. It does help with color and consistency, but lowers baseline quality and reduces the amount of possible change because of how it enforces the start image.
I focus on looping videos and my fix was to not use the Wan node pack. Use some default ksamplers or advance ksamplers. Also use a high/low checkpoint that is merged with lightspeed like this https://civitai.com/models/1981116?modelVersionId=2388548.
Was very annoying and took me a 3 weeks-1month to figure it out when Wan 2.2 came out so I can keep doing loops. Works the same for extending frame.