šļø this is a place for Minimax H3 Video workflows that are working in low-VRAM GPUs
Turbo LORA / Nodes from: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
or Lightx2v LORA: https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
Audio VAE from: https://huggingface.co/dummy9996/minimax_h3_vae_fp8/tree/main
Video VAE from: https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main
Uncensored TE from: https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/tree/main
Also you can Download Resources from any of these repos:
https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/tree/main
https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/tree/main
https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot/tree/main
and ...
Official ComfyUI Minimax H3: https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main
For live previews node get: https://huggingface.co/Kijai/MiniMax-H3-TAE
P.S.: you should learn how to use this model framework - it is really worth your tears and wasted hours of trial/errors ;)
Description
faster gens without attention / sampler tricks
FAQ
Comments (11)
How fast? On what hardware?
speed depends on steps + video duration, hardware < 12 GB Vram
@0l1v1aR0551Ā I see, thanks
@0l1v1aR0551Ā ok, but in your low VRAM hardware, what was the speed and the hardware?
@SantodanĀ 7 sec video - 22 min at 4 steps
NVIDIA GeForce RTX4080 with 12 GB VRAM and 32 GByte normal RAM 30 minutes for a 10 seconds video with sound (idealia, 8Steps)
@Mikel007Ā there is a possible reason for that - the kj node that feeds junks of model forward slows generation down / preventing VRAM overflow...
also some Comfyui settings like commandline --low-vram do the same effect - so, by optimizing your memory management you should have way faster gens of short vids (10-15 secs) on such nice hardware as you just mention!
Node 'ID #136' has no class_type. The workflow may be corrupted or a custom node is missing.: Node ID '#136'
update Comfy to the nightly (not stable) version for this to work!
Thank you very much for your work. However, with my 6GB+16GB laptop, I have tried all the Minimax low-VRAM version workflows and encountered an error with every one of them:
Error message: # ComfyUI Error Report ## Error Details - Node ID: 129 - Node Type: VAEDecode - Exception Type: RuntimeError - Exception Message: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument tensors in method wrapper_CUDA_cat)
I hope someone kind can help me figure out why I get an error every time the process reaches VAE video decoding.
go for INT4 safetensors wherever you can - look in all links of repos provided in this WF card - scan them regularly for new and / old INT convrot ones and test them on you computer = this is for low vram / ram issues;
but also - set Comfyui to nightly version !!! and update it several times a day ;)