The best way to upscale and optimize Hunyuan workflows is definitely something that will be debated for a long time, I've created a workflow that works best in my opinion.
Key Features:
TeaCache sampler - Allowing quicker video generation
Restore faces (ReActor) - Makes blurry and low quality faces look better (supports NSFW)
2nd pass including detail injection and latent upscaling for sharper video
3rd pass including upscaling and frame interpolation
This workflow comes pre loaded in my RunPod template
https://runpod.io/console/deploy?template=6uu8yd47do&ref=uyjfcrgy
Description
FAQ
Comments (7)
Thanks a lot, it's fun to check out different workflows, gonna use this for a few hours to see how it goes.
Works best for Runpod Only.
Takes 1266.29 seconds on a regular PC :D
it's about 3 minutes on first pass using default workflow settings on a 4090. a little less if i change it from bf16 model selected in workflow to the distilled fp8 one i usually use, but still 2 min 50 seconds. this is without changing any other settings. Sadly. it still has ghosting affects i get with other workflows making upscaling pointless. I keep finding time and time again i am better off just using higher resolution i can use on first pass. i don't get all hype on upscaling due to horrible image quality and ghosting. i tried default prompt a few times and after upscale there is ghosting halo looking affect around the hair. so again i prefer no upscaling. I'm tempted to not use comfyui for upscaling at all at this point and just find a 3rd party video upscaler.
Thanks for making this! Do you have any suggestions for what to tweak first to get this to run on 24gb vram? Or is it just meant for larger gpus on runpod?
Without changing the models I had to go all the way down to 256x384x45 just to get it to run. Using the fp8 checkpoint it's a bit better.
Why is a workflow installation never simple. Missing nodes, manager install missing nodes, still missing nodes
there's just no way I can install and get ReActorRestoreFace node working
