NODE FIX: The latest Comfy update broke the WaveSpeed (applyFBCacheOnModel) node. I have replaced it with an equivalent tea cache node, so you can retain same speed and quality outputs. This node should also be in Comfy Manager, so it will be easier to install and keep updated.

People keep asking me for my workflow, so I'm sharing it here. It honestly isn't anything special and I haven't bothered trying to optimize it much after I hit the sweet spot with it a few weeks ago. It is very likely that there are faster/more efficient ways to get similar results now.

The workflow used to have TorchCompile included in it, but it wasn't giving any perceived improvements and causing issues with my comfy, so I took it out. Feel free to add it back in if you would like to try to gain the speed increase and VRAM reduction with it.

The settings in this are what I produce my videos at. I sometimes upscale using the optional Simple Upscale, but a lot of the time it is quicker for me to take the original video output and pass it through Topaz. To be honest though, the initial video is decent enough to post even without upscaling. If you want to get fancy with it, change the simple upscale to an upscale with model.

On at 4090FE with 24gig of VRAM, I get the following:

89 frames = average of 11s/it, which equates to roughly 3 minutes to gen

121 frame = average of 14s/it, which equates to roughly 4 minutes to gen

145 frames = average of 22s/it, which equates to roughly 5.5 minutes to gen

Feel free to try lower frames with bf16 model too. I use fp8 to max out my frame count.

Mileage may vary for others. Enjoy!

NOTE: If you can't find the node title "applyFBCacheOnModel" in comfy-manager, you may have to install it manually from here: https://github.com/chengzeyi/Comfy-WaveSpeed?tab=readme-ov-file