Fastest possible workflow - use v2. better consistency with references - use v1. I have added a new version of the workflow (taken from banodoco discord - all credits go to the amazing community on there). more info in about this version. 5090 - 1mp 7s video generated in 60s. NOTE: I run my comfy on linux and the dlss node i've used might not work the same for you, you can use the windows native dlss5 nodes or skip it entirely if you want. ----------------------------------------------------------------------------------------------------------------- Here is the current workflow i am using to test my loras. Sharing it here, so people can critique it and maybe suggest improvements. on my 5090, it renders a 7 second 0.9 mp final 60fps video in ~60-80s. as you can see, when there's a lot of stuff going, fine detail gets very smeary.

Minimax H3 - Refmod + Reference - 2 Pass Latent Upscale Workflow

All credits go to the amazing community at the Banodoco discord, this was taken from there and adapted to my existing workflow. on my 5090 7 second video took only 60s at 1mp With this kind of speed, i didn't feel the need to add latent upscale - but you can if you want to.

Files with SHA256: 8972a187b59423f16862e5b39631290815fdbbf2714c4f3e53bfee1e47bdc375