My latest workflow for First Last Frame with Wan 2.2.
I've based this off of several other's workflows and simplified it down to the bare minimum I think is needed for high quality video generation. This includes:
First frame and last frame support to guide and control the scene and action
LoRA loading
Block swapping to help with VRAM
Interpolation and Upscaling for finishing touches
Two versions are available:
With TensorRT interpolation and upscaling (RTX50xx series cards only)
Standard interpolation and upscaling for all other video cards.
Description
FAQ
Comments (3)
You can use the WhiteRabbit BatchResize Lanczos node for fast upscaling. It has TorchLanc Upscaling: CUDA accelerated (10x faster), color correct and aspect correct.
Thanks for the tip. I think it'd be hard to beat TensorRT for speed on my 5090 as it is insanely fast compared to the standard RIFE VFI and standard upscaler, but I'll check out White Rabbit in general - looks like he has some nodes that I might find useful for batch processing.
Had trouble installing TensorRT (Yeah trouble installing a dependency, day ending in Y lol) so i tried your suggestion about the Whiterabbit nodes, and that sped up my workflow by over 2 minutes due to its acceleration. Absolutely incredible.
Though setting that cache offload part from every 10 frames to 0 (disabled) seemed to make it offload cache EVERY frame strange enough, but either setting didn't seem to have a HUGE impact on performance, given vram is already aggressively cleared in my workflow i'd just turn that off for a bit of a speedup since it's not trying to clear ram that's already cleared.
Pretty crazy what we can pull off with comfyui to save time with the right nodes lol.