Kijai

/

MiniMax-H3_comfy

Like 488

if you use another programs in you computer,that affects it,and on first use because of model loading it can be take longer than second use

Hello everyone. First of all, thank you Kijai, as always for your great work.

I tried the lightv2 loras figured that the 768p resized ligthv2 lora worked best for me.

I have 12GB vram and 64GB ram and in Comfy i use the "MiniMax H3 Mem Eff Sage Attention Patch" component.

I did some local animations with the resolution always being 640x640 and a variaty between 6 and 12 seconds, kept the same prompt for the same image and every run was randomly seeded and I pretty much didnt even see use the pc.

Now comes the confusing part.

Some 6 second videos took 4 minutes, others 20. Some 7 second videos took 5 minutes, other times 20.
Same prompt, steps and image, but different seeds had very different generation times. Some took three minutes others up to 12.

Am I just being unlucky and it reaches a close OOM and thats why it takes so long sometimes? (I mean it's ComfyUI behavior)

Anyone got similar results?

I'm having the same issue on a Krea2 with GGUF models (also 12 GB VRAM). It's a lack of video memory; at those moments (when it takes a very long time), the video memory literally hits the ceiling (I enabled the display of used memory in the system tray using MSI Afterburner). When it runs normally, it uses about 11,500 MB of memory. With the exact same settings, the behavior is strange—sometimes it uses more memory, sometimes less—so this might be a ComfyUI bug. I’m running MiniMaxH3 on the latest version of Comfy, with the new dependencies, and it runs stably (I’m using Unsloth’s pruned Q4 GGUF).

I'm having the same issue on a Krea2 with GGUF models (also 12 GB VRAM). It's a lack of video memory; at those moments (when it takes a very long time), the video memory literally hits the ceiling (I enabled the display of used memory in the system tray using MSI Afterburner). When it runs normally, it uses about 11,500 MB of memory. With the exact same settings, the behavior is strange—sometimes it uses more memory, sometimes less—so this might be a ComfyUI bug. I’m running MiniMaxH3 on the latest version of Comfy, with the new dependencies, and it runs stably (I’m using Unsloth’s pruned Q4 GGUF).

Yeah exactly. I was able to generate 15 seconds videos in 5 minutes yesterday. Today, it takes me 20 minutes to generate a 8 second video without me even touching the pc. I am eternaly confused at this point,

I'm having the same issue on a Krea2 with GGUF models (also 12 GB VRAM). It's a lack of video memory; at those moments (when it takes a very long time), the video memory literally hits the ceiling (I enabled the display of used memory in the system tray using MSI Afterburner). When it runs normally, it uses about 11,500 MB of memory. With the exact same settings, the behavior is strange—sometimes it uses more memory, sometimes less—so this might be a ComfyUI bug. I’m running MiniMaxH3 on the latest version of Comfy, with the new dependencies, and it runs stably (I’m using Unsloth’s pruned Q4 GGUF).

Yeah exactly. I was able to generate 15 seconds videos in 5 minutes yesterday. Today, it takes me 20 minutes to generate a 8 second video without me even touching the pc. I am eternaly confused at this point,
at what resolution and what steps?

I figured it out. It is the image. The clearer the image, such as zoomed in, the faster it understands what you want him to do...
Edit: Never mind, it actually is the prompt. The longer you make it and the more details you add will dramatically increase the gen time.

at what resolution and what steps?
4 steps with the turbo lora and 640x640

at what resolution and what steps?
4 steps with the turbo lora and 640x640

hmm,good you figured it out,i can say minimax h3 is so slow,im generating 480p videos 5 seconds and 20 steps in 8 minutes,786p with 20 steps takes almost 21 minutes,and im using qwen 3 vl 4b with clipproj too for give more vram for minimax

I have made many more tests and I have figured out that it was neither the prompt length nor the detailed image.
I am using JoyCaption which gives a description of the image. The option to save the caption inside the VRAM allocated a bunch of it. upon reaching the "Requested to load MiniMaxH3" section, there was 0.00MB available, resulting in weird comfyui behavior.

Turns out looking at the information inside the cmd can be quite helpful.

I downloaded the latest version of ComfyUI, and with it, I can render at a resolution of 864x864 (0.7 MP) in 15 seconds, all with 12 GB of VRAM. The model is “minimax_h3_fl2va_pruned-Q4_K.gguf” by Unsloth, and the loader is this one: https://github.com/molbal/ComfyUI-GGUF
Clip model: “qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf” by Rebel.
https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/tree/main
https://huggingface.co/unsloth/MiniMax-H3-GGUF/tree/main

I downloaded the latest version of ComfyUI, and with it, I can render at a resolution of 864x864 (0.7 MP) in 15 seconds, all with 12 GB of VRAM. The model is “minimax_h3_fl2va_pruned-Q4_K.gguf” by Unsloth, and the loader is this one: https://github.com/molbal/ComfyUI-GGUF
Clip model: “qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf” by Rebel.
https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/tree/main
https://huggingface.co/unsloth/MiniMax-H3-GGUF/tree/main

0.7 mp in 15 seconds?you are faster than api you must have like 3tb/s gpu