Update 8/3 - Since Wan 2.2 High and Low combination lora sets are starting to be pushed out, nodes have been added to support both High and Low only loras applied to the appropriate model, plus a section for 2.1 loras (or if you use low-noise only loras) that you want applied to both passes. Also merge_loras on scaled models now works, make sure to update your WanVideoWrapper node suite!

My somewhat condensed workflow, meant for 24Gb+ VRAM cards. Mostly set up for a 5090, but can easily be used with a 3090 or 4090 by upping the block swap to fit. Lots of workflows are out there for folks with less powerful GPU's, but if you have the power, why not use it?

Also you should have at least 64Gb of system RAM as well, to handle model/lora offloading. ComfyUI's poor RAM clearing may require you to close the terminal and re-launch Comfy from time to time, if you're doing a lot of generations and swapping out loras. I've had it eventually run out my 96GB of RAM. And be careful, WAN 2.2 will memory creep faster than 2.1 did, keep an eye out so you don't lock up your system!

Also you'll need a higher block swap than what 2.1 required.

Make sure both your comfyui install and wanvideowrapper nodes are up to date.

Lightx2v lora-

https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_14B_T2V_cfg_step_distill_lora_adaptive_rank_quantile_0.15_bf16.safetensors

Kijai's WAN 2.2 Quants (You'll need both HIGH and LOW models)-

https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/T2V

Outputs will be in a dated WAN2_2_T2V folder, and the final frame saved as well, in case you want to play around with extending in I2V.

There are 2 quality of life options next to the prompt box, one is "Clear Cache After Generation", which will unload the model after a render. Useful if you notice comfy not letting go of your VRAM between gens. Sadly will not clear system RAM.

The other option is to turn off frame interpolation, since once the workflow hits that node it's stuck until it processes. You can do a generation with this off, and then if you like the output just turn it on and it'll interpolate the video you just rendered. But if you change anything before you toggle that on, it'll re-render the entire video.

On a 5090, with the model already loaded it puts out high quality gens at 81 frames at 1024x640 in about 87 seconds, and interpolation brings that up to 129 seconds.