It takes 6 minutes for minimax-h3 to generate a 5-second, 480p video on DGX SPARK; how long does it take for yours?
i2v: 5s/0.2M, takes 174s
t2v: 5s/0.2M, takes 143s
r2v: 5s/0.2M 2 reference pics, takes 215s
I am about 235s for 5 min video using megalpixel 0.6 1:1 aspect ratio which equals to 768x768.
it took 432s for this 10min video using easycache & patch sage attention kj nodes.
You are generating 5 min videos? can you share your workflow?
If you are using comfy, go to templates and you should see the 3 available:
download models from here https://huggingface.co/Comfy-Org/MiniMax-H3
if you want to use easycache & sage attention, right click on image to video minimax h3 and unpack subgraph to show you all the nodes.
To add easycache and sage attention, use download KJNodes and easycache is native to comfy, attach them together then model to easycache and attached KJ to basic scheduler and basic guider
How much of a speedup does sage attention offer here?
Use this for dgx spark, after install and rebooted. Go to the folder /custome_nodes/ComfyUI-H3-SolEngine/workflows and use one of the three like h3-fullstack.json and it should already be setup. This improves the workflow a bit, I’ve tried different addons and this workflow work a bit better for me.
I also recommend using the turbo lora.
larryvrh/MiniMax-H3-Turbo-Lora · Hugging Face
git clone or search MiniMax-H3 Turbo in the manager.

