Trying out MiniMax H3 in ComfyUI

MiniMax H3 has finally been released as open source.
Although there are license restrictions based on certain regions, it seems it can be used in Japan without any issues for now. (Please check the license for details.)

When I tried it with paid nodes, it felt great with high consistency, but the question is how much of that can be achieved locally.

The workflow is based on the following.

Environment

OS: Windows 11
GPU: GeForce RTX 4090
CPU: i9-13900KF
Memory: 128GB

Setup Procedure

Please update ComfyUI to the latest version.

Also, since it won't run as-is with my machine's specs, I added the "--reserve-vram 6.0" option to the ComfyUI startup bat file as follows.

ComfyUI_windows_portable\run_nvidia_gpu.bat

Before change

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build

After change

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --reserve-vram 6.0

--reserve-vram 6.0: The amount of VRAM (in GB) that ComfyUI will "leave empty without using" when loading models. This is the space reserved for the OS and other applications.

* It is recommended to create a separate run_nvidia_gpu.bat file rather than using the existing one.

Models

minimax_h3_fl2va_pruned_int8_convrot.safetensors and minimax_h3_ref2va_pruned_int8_convrot.safetensors downloaded and placed in "ComfyUI\models\diffusion_models"

qwen3vl_32b_minimax_h3_int8_convrot.safetensors downloaded and placed in "ComfyUI\models\text_encoders"
(For 50X0 series, qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors is also acceptable)

minimax_h3_audio_vae_fp32.safetensors and minimax_h3_video_vae_fp16.safetensors downloaded and placed in "ComfyUI\models\vae"

Workflow (I2V)

The workflow is as follows.

The generation results are as follows.
Generation time: 10 minutes.
CPU memory: 90GB
GPU memory: 28GB (21GB VRAM + 7GB shared memory)

Workflow (I2V)

The workflow is as follows.

The generation results are as follows.
Generation time: 11 minutes.
CPU memory: 90GB
GPU memory: 29GB (22GB VRAM + 7GB shared memory)

Impressions

The generation results themselves are quite high quality.
However, the required memory is simply huge.
As it stands, it is completely aimed at the 5090.

If you increase the resolution or frame count, the required memory increases proportionally, so it is quite difficult even with a 4090.

That said, the number of steps is around 20, so I expect mechanisms to reduce this, as well as further quantized models, to be released, which I am looking forward to.

I performed additional verification, and you can expect an improvement in generation speed by using CUDA 13.
Please check it out if you are interested.