Official ComfyUI workflows with fl2va int8 version of the Minimax H3 model generating 8 sec 720p video on rtx4080 16GB + shared RAM: 32GB
with no optimizations generates for 25 min
with FirstBlockCache reduces generation to 15 min - lossless quality
just works or set mode to H3 Aggressive for even more speed
with Turbo Lora reduces generation to 6 min - with latest version the quality loss is minimal on 6 steps
Other changes from official workflow:
replaced Load Image node with Image Crop for convenience to make the image in the video aspect ratio to prevent stretching
replaced Save Video with VHS version of it to improve the codec quality of the video file
added notification node to alert when generation is finished and display generation time
added Comfy Kitchen Attention node for optimization
Updates:
changed all models to int8 - better quality
updated turbo lora - better quality and speed
Description
FAQ
Comments (5)
Ok, just tried it. I have a 5090 and 96gb ram. I typically do 1024 x 1024 @ 10 seconds. Normally this has been taking about 24 minutes. With this workflow and the cache node it reduced generation time to about 13:30. Well done, thank you <3
PS: Implementing the workflow was easy, no issues. Just had to learn to use the crop/image node.
thanks for the feedback and the buzz tip
crop image node is more convenient than the official load image node because you need to make the image in the video aspect ratio to prevent stretching, works best by just drag and drop image into it and choose aspect ratio and position
cache is really good lossless solution that just works
turbo version can greatly reduce generation time but you have to play around with the configuration to reduce the quality loss which in the end will take more time for testing, maybe in the future when it improves can be more usable
@goonjin Yeah, I tried one of the turbo loras yesterday, the quality was pretty bad. They just need more time, they will absolutely get there.
I also realized today that you can do text to image with your workflow, you just need to bypass the load image node and disconnect it. Worked fine.
@jbear2447 the cache node can be used for all Minimax H3 workflows, just connect Load Diffusion Model to it, then connect it to Basic Scheduler and Basic Guider
https://github.com/duckyshell/ComfyUI-MiniMaxH3-FirstBlockCache#connection
Thanks for the feedback, I have a 5090 as well, and 24 minutes for a 10 seconds video was just too much for my taste.. 13 minutes, for a good quality video, I can live with.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.