MiniMax H3
Built on the default MiniMax H3 workflow from ComfyUI with the added 'Spectrum Apply MiniMax H3' node (credit to the author xmarre )
"Spectrum replaces selected expensive transformer evaluations with spectral feature forecasts while preserving MiniMax H3’s native audio/video output and reconstruction path." - xmarreOn my local setup, I was able to generate videos 3x faster with a RTX Pro 6000:
5 sec video in 16 seconds
10 sec video in 120 seconds
Model Links
node
ComfyUI-Spectrum-MiniMax-H3 - https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3) (MUCH faster generation time, ~3x fast, credit to the author xmarre)
vae
minimax_h3_video_vae_fp16.safetensors - https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
minimax_h3_audio_vae_fp32.safetensors - https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
diffusion_models
minimax_h3_fl2va_pruned_int8_convrot.safetensors - https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
text_encoders
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors - https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Description
FAQ
Comments (17)
Hey great work, can it help my 3090 output faster, even after sage attention?
it does what is says, on 3090 I got 0.4mp at90sec, but got body horrors as well. I did not use any first last frame though. could be that, still testing
Isn't any faster at all on my 5070.
It works like a charm!
This runs slower than the stock workflow on my RTX 5060 Ti 16GB.
i9 14900k and 5090 btw, gg
IndexError: tuple index out of range 2026-08-04T15:54:04.925190 - [32m[INFO][0m [32mPrompt executed in 150.70 seconds[0m 2026-08-04T15:54:05.411849 - [MultiGPU_Memory_Monitor] CPU usage (89.0%) exceeds threshold (85.0%) 2026-08-04T15:54:05.421051 - [MultiGPU_Memory_Management] Triggering PromptExecutor cache reset. Reason: cpu_threshold_exceeded
this worked great indeed
Shaves off roughly 40 seconds from a 10s video for me ony my RTX 6000 (Max-q version, using sage attention). Down from roughly 145s to 105s. From 55s to 43s on 5s videos. So a nice improvement, but not quite the claimed 3x. But still quite nice!
Those times include vae decoding/saving to disk + maybe a couple seconds from processing the prompt, if it's just the generation bit before that, it's 87s vs 120s for 10s, and 33s vs 48s for 5s with this node added.
A 5 second video without the node takes ~67s for me, with node ~16s, so actually faster than 3x. I also use sage-attn. Are you on windows perhaps?
@ultimo_intento Yes, windows 10. Not that it matters that much to me, as long as 10s+ is faster, and that seems to match what you're getting. Weird that your 5s takes so much longer compared to mine, without the node, and so much faster with the node. this is with 0.4 res?
@PahviKahvi ah, makes sense. You probably don't have the latest sage-attn and cuda runs faster on Linux also
Yeah, I think my 67s quote is with 0.5mp
@ultimo_intento Yes, I don't have sageattn 3 installed. Might have to consider setting up a dual boot I guess, if there's notable speed differences, what res is your 120s with node 10s gen done on?
@ultimo_intento Installed SA3, shaved off another 10s from the 10s video gen, but it does slightly lessen the image quality more, than SA2 from what I can tell. Not a massive difference, but noticeable at least at 0.4.
Ref to Video?
I'm not sure what you're asking? You can drop in the node on the img and vid ref workflows just the same.
using a 5090, 10s, 0.6mp, with added rtx super scaler default 2x/ultra res. Went from 208s to 178s. I really hope that turbo lora pans out, this is already fast af.