<h3 id="important-warning">⚠️ IMPORTANT WARNING ⚠️</h3><p>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.</p><h3 id="what-it-does">🛠️ What It Does</h3><p>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.</p><h3 id="how-to-start">🚀 How to Start</h3><ol><li><p><strong>Install Dependencies:</strong> You must install the MacH3 Metal node pack and the h3.c engine (links are below).</p></li><li><p><strong>Run Workflow:</strong> Start with <code>macH3_metal_all_in_one.json</code>. It opens ready for text to video with a short prompt, so you can just press Run.</p></li><li><p><strong>First Load:</strong> The node pack copies its sample pictures and clip into ComfyUI's input folder on the first run.</p></li></ol><h3 id="usage-and-settings">⚙️ Usage &amp; Settings</h3><ul><li><p><strong>Modes:</strong> Select a mode on the Take node (T2V, I2V, FL2V, or R2V with up to 3 references).</p></li><li><p><strong>Continuation:</strong> 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.</p></li><li><p><strong>Settings:</strong> You can adjust the size in megapixels and the lane: Quality (12 steps) or Speed (using the Turbo LoRA at 6 steps, same weights).</p></li></ul><h3 id="performance-notes-(m5-pro-48-gb)">⏱️ Performance Notes (M5 Pro, 48 GB)</h3><p>On my M5 Pro (48 GB), here are the times for a 3 second clip:</p><ul><li><p>0.5 MP takes: 3:25 on Speed, 6:14 on Quality.</p></li><li><p>1.0 MP takes: 7:05 on Speed, 12:15 on Quality.</p></li></ul><p>ComfyUI peaked under 10 GB of memory on all four tests because the 18 GB of int8 weights stream from the SSD.</p><h3 id="system-requirements">🍎 System Requirements</h3><ul><li><p>Apple silicon Mac (Tested on 48 GB M5 Pro).</p></li><li><p>ffmpeg.</p></li><li><p>MiniMax H3 weights.</p></li><li><p>h3.c engine built with <code>make libh3.dylib</code>.</p></li><li><p>The node pack.</p></li></ul><p><em>Note: 24 GB+ on M1 or newer should work, but this is untested.</em></p><p>The first run builds an 18 GB int8 cache. Full steps are in the repo README.</p><h3 id="credits-and-source">🔗 Credits &amp; Source</h3><ul><li><p>MiniMax H3 by MiniMax</p></li><li><p>h3.c by antirez: <a target="_blank" rel="ugc" href="https://github.com/antirez/h3.c">https://github.com/antirez/h3.c</a> (my fork with the library build: <a target="_blank" rel="ugc" href="https://github.com/Bambushu/h3.c">https://github.com/Bambushu/h3.c</a>)</p></li><li><p>Turbo LoRA by larryvrh: <a target="_blank" rel="ugc" href="https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora">https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora</a></p></li></ul><p>Source, MIT license, and updates are available here: <a target="_blank" rel="ugc" href="https://github.com/Bambushu/ComfyUI-MacH3-Metal">https://github.com/Bambushu/ComfyUI-MacH3-Metal</a></p><p>I entered this in the Open Source Higgsfield challenge. #OpenSourceHiggsfield</p>

MacH3 Metal: MiniMax H3 for Apple Silicon (M1+, 24GB+ RAM)

<p>First release. All four modes (T2V, I2V, FL2V, R2V) in one workflow, continuation across takes, Quality and Speed lanes, plus an App Mode version.</p>

Files with SHA256: 1a15d7ec359c10a8039e75543f430c174273d1f887470a08faeaa50b731f7c7d