⚠️ IMPORTANT WARNING ⚠️

This workflow requires the MacH3 Metal node pack and the h3.c engine installed. It will not run on stock ComfyUI or on NVIDIA cards.

🛠️ What It Does

I built this because I wanted MiniMax H3 running natively on my Mac without needing the PyTorch route. This is a node pack plus workflows, based on h3.c, a native C and Metal engine for H3. One pass gives you both the picture and its own sound.

🚀 How to Start

1. Install Dependencies: You must install the MacH3 Metal node pack and the h3.c engine (links are below).

2. Run Workflow: Start with macH3_metal_all_in_one.json. It opens ready for text to video with a short prompt, so you can just press Run.

3. First Load: The node pack copies its sample pictures and clip into ComfyUI's input folder on the first run.

⚙️ Usage & Settings

• Modes: Select a mode on the Take node (T2V, I2V, FL2V, or R2V with up to 3 references).

• Continuation: To continue a shot, set Continue to "last take" in Settings, write what happens next, run again, and the Shot node joins the takes into one clip.

• Settings: You can adjust the size in megapixels and the lane: Quality (12 steps) or Speed (using the Turbo LoRA at 6 steps, same weights).

⏱️ Performance Notes (M5 Pro, 48 GB)

On my M5 Pro (48 GB), here are the times for a 3 second clip:

• 0.5 MP takes: 3:25 on Speed, 6:14 on Quality.

• 1.0 MP takes: 7:05 on Speed, 12:15 on Quality.

ComfyUI peaked under 10 GB of memory on all four tests because the 18 GB of int8 weights stream from the SSD.

🍎 System Requirements

• Apple silicon Mac (Tested on 48 GB M5 Pro).

• ffmpeg.

• MiniMax H3 weights.

• h3.c engine built with make libh3.dylib.

• The node pack.

Note: 24 GB+ on M1 or newer should work, but this is untested.

The first run builds an 18 GB int8 cache. Full steps are in the repo README.

🔗 Credits & Source

• MiniMax H3 by MiniMax

• h3.c by antirez: https://github.com/antirez/h3.c (my fork with the library build: https://github.com/Bambushu/h3.c)

• Turbo LoRA by larryvrh: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora

Source, MIT license, and updates are available here: https://github.com/Bambushu/ComfyUI-MacH3-Metal

I entered this in the Open Source Higgsfield challenge. #OpenSourceHiggsfield