SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

How to Install MiniMax H3: Works with 8GB VRAM + 16GB RAM ~I2V (Image to Video)~

ワンスタ技術部 | AI LAB

What is MiniMax H3?

MiniMax H3 is an open-weight model released by MiniMax on August 3, 2026. It is a multimodal generative AI model that can be used for free within the scope of its license. Specifically, you can generate videos by inputting text, images, video, and audio. There are several models available depending on the purpose, and you should use the appropriate model for each task.

T2V (Text to Video)

This is a method familiar to anyone who has generated images. You write a text prompt and use the concepts learned by the MiniMax H3 model to generate a video based on it. If you are already able to generate still images with tools like Anima or Krea2, you should use the I2V method introduced in this article. Since most people prefer I2V over T2V, this article will not cover T2V.

I2V (Image to Video)

This method generates videos by using an existing image, describing the outline of the image in a prompt, and writing multiple optional cut prompts. If you have used Wan2.2, you may know that it does not generate audio. While LTX2.2 can generate audio, Japanese speech synthesis is also possible with MiniMax H3.

Ref2V (Reference to Video)

This is the most complex model of MiniMax H3, known as the Omni-Reference model. It allows you to generate videos by combining reference materials such as images, videos, and audio. Specifically, you can define character designs with images, extract movement by referencing videos, and define zero-shot speech synthesis simultaneously, outputting it as a single integrated video. Although complex, it is a very powerful model capable of creating videos equivalent to the now-discontinued Sora2. I will only introduce it here, but since I have already finished verifying the operation of Ref2V, I would like to cover it in the next article.

Flow of this article

It is common to use ComfyUI to run MiniMax H3. If you have read the previous article and are practicing, you can generate videos simply by updating your existing ComfyUI to the latest version and installing the MiniMax H3 model. While it is possible to install MiniMax H3 after following the Anima installation steps from the previous article, I will start from the ComfyUI setup process for those who are new to ComfyUI.

  1. Operating Environment

  2. Installing Python and Git for Windows

  3. Installing ComfyUI

  4. How to update ComfyUI for those who read the previous article

  5. Installing MiniMax H3

  6. Generating videos with I2V

Since settings are mainly performed via the command line, if you encounter parts you don't understand, please input those commands into ChatGPT to confirm what they do before executing them. If you set up ComfyUI using this method, you will need to perform future updates via the command line as well. This knowledge will be useful when dealing with errors or other issues.

Operating Environment

  • It runs on Windows 11 (it is almost the same for Linux, but GPU drivers and CUDA settings are required).

  • As long as the CPU is supported by Windows 11, no particularly high specs are required.

  • For GeForce RTX 2000 series and later (with 8GB or more VRAM), it is possible to run it even with 8GB of VRAM using ComfyUI startup options.

  • Make sure your GPU drivers are up to date. We will use PyTorch for CUDA 13.0.

  • While 64GB of main memory is ideal, it will work with 16GB. In this case, the --disable-pinned-memory option is required.

  • Ensure you have at least 100GB of free space on your SSD. The MiniMax H3 models are massive, with the two I2V/Ref2V models alone taking up about 60GB.

  • ComfyUI is used by connecting locally via a web browser. Chrome, Edge, or Brave are recommended.

Installing Python and Git for Windows

We will use Python version 3.13.15 because it is the version recommended in the official Comfy.Org documentation, and because binary-level compatibility when installing pre-built binaries (Python packages) may shift to the 3.13 series in the future. If there is a patch-level update, it is good practice to update if a Windows build is provided. This reduces risk if there are security fixes. However, if you are replacing it in an active environment, you will need to rebuild the venv from scratch. Especially if you do not add the '--listen' option to ComfyUI and do not intend to use it over the internet, you may choose not to update it unnecessarily.

Install the latest version of Git for Windows. During installation, check the option to display a popup if there is an update. Since Git is a tool that downloads source code directly over the internet, you should always apply security fixes.

Once downloaded, you can install either one first. However, make sure to set both Python and Git to your system's PATH variable. Otherwise, you will have to run commands like `set PATH=C:\...\Python\bin;%PATH%` in the terminal every time. Git for Windows seems to be added to the PATH environment variable by default, so you can proceed as is. For the editor selection, choose VSCode if you have it installed. If not, you can specify Notepad. Since this article does not involve much use of an editor, Notepad is fine.

Python setup
Git for Windows update/adding to terminal
Editor to use with Git
Using Git in the default Windows terminal

Installing ComfyUI

This time, we will install ComfyUI v0.30.2 via manual installation. If you already have ComfyUI set up from a previous article, the update method will be described in the next section.

Installing ComfyUI

This task uses the Windows command line. Press Windows key + R, type 'cmd', and open the terminal.

Next, please copy and paste the following command lines one by one and execute them. Execute by pressing the Enter/Return key after pasting. While it will work if you copy and paste the entire text, it is better to execute it line by line so that it is easier to understand if any errors occur. You can paste command lines by right-clicking in the terminal.

c:
mkdir C:\Software
cd C:\Software
git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
py -3.13 -m venv venv
.\venv\Scripts\activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
pip install comfyui-manager

The reason for installing PyTorch first is that installing from requirements.txt may sometimes result in the CPU version being installed.

The issue here is that with a clean ComfyUI installation, you need to activate the Python virtual environment and run the script with `python` every time you start it. Therefore, let's create a batch file to make it easy to launch.

Press the Windows key + R, type `notepad` to open Notepad, and paste the content below that matches your environment.

  • For environments with 16GB VRAM and 64GB main memory

@echo off
call .\venv\Scripts\activate.bat
python -s main.py --enable-manager

  • For environments with 12GB VRAM

@echo off
call .\venv\Scripts\activate.bat
python -s main.py --lowvram --enable-manager

  • For environments with 8GB VRAM or low main memory
    *If you want to run it in an 8GB VRAM environment or if you have less than 64GB of main memory, please add --disable-pinned-memory after --enable-manager. This option prevents parts of the model that don't fit in VRAM from being placed in main memory, instead loading them from the SSD as needed. It runs in a practical amount of time even on an RTX 5060 (8GB VRAM) + 16GB main memory setup; generating a 5-second video at 0.4Mp (864x480 resolution) takes approximately 7 to 8 minutes.

@echo off
call .\venv\Scripts\activate.bat
python -s main.py --lowvram --enable-manager --disable-pinned-memory

Save this text as `C:\Software\ComfyUI\run.bat`. If the extension becomes `.txt`, go to `C:\Software\ComfyUI` in File Explorer and rename `run.bat.txt` to `run.bat`. Since it is tedious to navigate this path in Explorer every time, you may want to create a shortcut on your desktop.

How to update ComfyUI for those who read the previous article

If you encounter errors while following these steps, try selecting and pasting the terminal text into ChatGPT to investigate the cause. Note that this article uses Python 3.13.15; if you are already using 3.12.10, be aware that you will need to rebuild the venv. If that is too much trouble, you can rename `C:\Software\ComfyUI` to something like `ComfyUI-v0.27.0` and start from the section above.

c:
cd c:\Software\ComfyUI
.\venv\Scripts\activate
git switch master
git fetch --tags
git tag
git branch 202608-local
git commit -a -m "requirements.txt changed."
git checkout v0.30.2
pip install -U -r requirements.txt
pip freeze > requirements-backup-202608.txt

The `git commit -a -m "requirements.txt changed."` part is necessary if you commented out the torch-related lines in `requirements.txt`. It is not necessary if you have not modified the ComfyUI git repository.

Installing MiniMax H3

When you open the MiniMax H3 template in ComfyUI, it will automatically download the necessary models, but it is more reliable to download and place them manually in a stable environment. Although it may have been fixed by now, downloads from within ComfyUI sometimes cannot be resumed if they fail.

The official MiniMax H3 models are intended for use with diffusers and are not designed to be handled by ComfyUI. However, Comfy.Org provides models adapted for ComfyUI, so we will download and use those.

・Placement location

C:\Software\ComfyUI\models\diffusion_models\MiniMaxH3

*The MiniMaxH3 folder does not exist, so please create it using File Explorer.

・Placement location

C:\Software\ComfyUI\models\text_encoders\MiniMaxH3

*The MiniMaxH3 folder does not exist, so please create it using File Explorer.

・Installation Location

C:\Software\ComfyUI\models\vae\MiniMaxH3

Note: The MiniMaxH3 folder does not exist, so please create it using File Explorer.

Generating video with I2V

Once you have reached this point, let's launch ComfyUI. Double-click and run the C:\Software\ComfyUI\run.bat file you created earlier.

[INFO] To see the GUI go to: http://127.0.0.1:8188

If the local URL is displayed in the console, it is a success. Ctrl+Click the URL to open ComfyUI in your web browser.

Opening the template

Since building the workflow manually is difficult, let's call up and use the MiniMax H3 template. From the ComfyUI toolbar, click Template and enter MiniMax in the search box. Then, a template called MiniMax H3 Image to Video will appear; click it.

ComfyUI toolbar
MiniMax H3 template
Template workflow

Since the model paths here differ from what the template expects, switch them as follows:

  • `unet_name` → `MiniMaxH3/minimax_h3_fl2va_pruned_int8_convrot.safetensors`

  • `clip_name` → `MiniMaxH3/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`

  • `vae_name` → `MiniMaxH3/minimax_h3_video_vae_fp16.safetensors`

  • `audio_vae` → `MiniMaxH3/minimax_h3_audio_vae_fp32.safetensors`

Changing model paths

Setting the image to use for I2V

Next, set the image to use for I2V. Drag and drop the image you want to use from File Explorer into the node labeled Load Image on the ComfyUI workflow. The screen will then look like the one below. The image itself can be one generated by Anima or ChatGPT. I have also attached an example image below, so feel free to download and set it.

Result of dragging and dropping into image loading
For download (please right-click to save)

Prompt Settings

Let's modify the description slightly from the template so we can animate this image.

Editorial tech product film. The animal woman from <Picture 1> in its original
scene: A woman wearing miko outfit stands in the foreground, with a forest of fresh green foliage spreading out and torii gate in the background.
SHOT 1: The woman moves her upper body slightly and waves her right hand.
SHOT 2: The woman is doing knee bends and stretching exercises.
SHOT 3: The woman runs toward the foreground, and the scene opens up to reveal only a forest.
Audio: A woman says, "キツネはコンコン鳴かないんですけどね"

Setting the Image Aspect Ratio

Since it is a portrait source image, select 3:4 (Portrait Standard).

Aspect Ratio Settings

Video Generation

Click the light blue Queue Prompt button at the top of ComfyUI. This will start the video generation. Once generation is complete, you can preview it in ComfyUI. Press the play button at the bottom left of the Save Video node.

Executing the Workflow

The generated video will be output under C:\Software\ComfyUI\outputs\Video\. Below is an example of the generated result.

Afterword

This completes the basic I2V video generation. Looking at the template prompt, it is structured as follows:

  1. Specify the video style, describe the features of the subject in the image, and recognize the character with `<Picture 1>`

  2. Describe the character's appearance and background situation with `scene:`

  3. Describe the character's actions for each shot with `SHOT n:`

  4. Specify the Japanese text for the character to speak with `Audio:`

It is structured in this way.

If you want to learn more about prompt writing, please read the official MiniMax H3 documentation.

Great job on installing MiniMax H3.

いいなと思ったら応援しよう！