MiniMax H3 Director 工作流界面

Configuration: Windows, RTX 5080 16GB video memory, 32GB DDR4, ComfyUI + MiniMax H3 Director.

First, let's talk about whether it can run: 16GB video memory is sufficient, but you must use quantized weights and avoid full precision. 32GB of memory is enough, and it's recommended to set the page file to 48-64GB. The disk should have at least 60GB of free space - the total weight is 63GB, and if you only run text-to-video, it's 42GB.

You don't need to tweak the parameters yourself, start with this set:

864×480 / 24fps / 124 frames (about 5.17 seconds) / 25 steps / res_multistep / simple / CFG 1.0 / video shift 12 / audio shift 3

I. Downloading Weights

Official repository: Comfy-Org/MiniMax-H3

If you only want to run text-to-video for now, you only need fl2va + text encoder + two VAEs. ref2va is only needed when doing reference images/videos, and can be downloaded later.

By the way: don't write Hugging Face tokens into your workflow or post them in articles.

II. Installation

1. Drivers and directories. Install the RTX 5080 drivers, and run nvidia-smi in the terminal to confirm that the card is recognized. I'm using the directory D:\minimax-h3\ComfyUI, and I've created the models/diffusion_models, models/text_encoders, models/vae, and custom_nodes directories.

2. Let ComfyUI run empty once. Install it according to the official Windows method, and then:

cd D:\minimax-h3\ComfyUI
.\.venv\Scripts\python.exe main.py --listen 127.0.0.1 --port 8188

If the browser can open the page, it means the base is fine, exit.

3. Install Director. Search for ComfyUI_MiniMaxH3_Director in ComfyUI Manager, or manually:

cd D:\minimax-h3
git clone https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director.git .\ComfyUI\custom_nodes\ComfyUI_MiniMaxH3_Director
.\ComfyUI\.venv\Scripts\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI_MiniMaxH3_Director\requirements.txt

4. Assign weights. Copy according to the target directory above, do not change the file name, and then restart ComfyUI.

5. Import workflow. Download T2V workflow JSON and drag it in. Select fl2va for UNET; select Qwen3-VL for CLIP, and choose the minimax type; the two VAEs correspond to video and audio respectively, do not connect them incorrectly.

6. Fill in parameters. It's the first group, connect CreateVideo and SaveVideo properly.

7. Run the first one. Use this short prompt, click Queue Prompt, and then confirm that the exported MP4 has both pictures and sound:

Realistic live-action cinematic shot of a quiet modern room at sunrise, soft natural light, subtle room ambience, no text, no watermark, no non-diegetic music.

Three, Form a Complete Story

Control each paragraph to be around 5 seconds, and only arrange one main action per paragraph, as trying to do too much will lead to failure.

The method is: after generating the previous paragraph, extract the last frame and use it as the first frame of the next I2V paragraph, referencing it with <Picture 1> in the prompt. The character's clothing, weather, vehicle color, and left-right positional relationships must be repeatedly locked in each paragraph's prompt, otherwise, it will drift away.

Subtitles and BGM are left for post-production: subtitles are compressed in post-production, BGM is uniformly added, and cross-fading is done between clips.

Resources

Official H3 Documentation　|　H3 Director GitHub　|　Chinese README