✦ Enzino H3 Simple · Full edition 🔒

MiniMax H3 is a 33B omni-modal video model that makes the sound and the picture in the same pass, and its ComfyUI setup is not small: a 21 GB diffusion model, a 32B Qwen3-VL text encoder, two VAEs, a Turbo LoRA. Open a working H3 graph and count the wires, and you'll see why someone wanted a node that says "give me a picture and a sentence, I'll handle the rest." That's ✦ Enzino H3 Simple.

It's also Full-edition only, which is the first thing to know about it.

The node you're looking at may be a lock

The pack ships two editions of the same code. Free is the six-node WF chain (Models → LoRAs → Prompt → Render → Enhance → Output). The paid edition ($11.99 on Ko-Fi) adds Simple, the big Studio node, and the production tools. In the free build EnzinoH3Simple still registers - but as a stub: a 🔒 card with no inputs that stops the queue with a plain-English message pointing at the Ko-Fi page and at workflows/enzino_h3_wf.json.

That's decent design, actually. Your workflow doesn't open with a red "missing node" rectangle and a manager hunt; you get a card that explains itself. But if you installed from GitHub and wondered why the sparkle node does nothing, that's why. Nothing is broken and nothing is missing.

What it actually is under the card

Worth knowing before you spend the twelve dollars: EnzinoH3Simple is literally a subclass of the Studio node. In the source, the only thing the Simple class changes is its description string; the engine, sampling path, memory handling and decode are the same object, and the compact face is the JavaScript layer.

So Simple, Studio and the WF chain produce the same video at the same speed. What you're buying is ergonomics - how many settings you have to look at - plus the Full-only extras like the Krea 2 second pass, story mode and 4K. Want "clip from this photo, five seconds, 16:9"? Simple is the shortest path. Want to know what the sampler is doing? The WF nodes are the same quality with the wiring visible.

What you set, and what it decides for you

The card is deliberately short: a mode (text, image, first + last, references), the picture slot or slots, the prompt, duration, format, quality. Models and LoRAs on a second tab.

Everything else runs on the pack's defaults, and they're the right ones: attention auto-picks the fastest backend you actually have installed (SageAttention before Flash before PyTorch), each transformer block's feed-forward gets chunked automatically on cards under 20 GB, the prompt cache skips re-encoding through the 32B text encoder when only the seed changes, and the patched model is reused between runs. On a 16 GB card those defaults are the difference between a clip and an out-of-memory error, and you set none of them.

It's an output node, so there's nothing to wire downstream - it hands back the video, the frames and the audio, with a player in the node so you can hear the stereo track without leaving the canvas. Duration snaps to H3's 17k+5 frame grid at 24 fps, so "five seconds" is really 124 frames.

Install

ComfyUI Manager, search the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/byenzinoai/ComfyUI-Enzino-H3-Studio
# restart ComfyUI, then Ctrl+F5 in the browser

There is no pip install step - pyproject.toml declares an empty dependency list on purpose, because the pack drives ComfyUI's own MiniMax H3 implementation instead of shipping model code. What it does need is a recent ComfyUI with native H3 support, and the weights.

Those are the real install, roughly 40 GB: the keyframes model (minimax_h3_fl2va_pruned_int8_convrot.safetensors, 21 GB, or the 16 GB W6A8 file on a 16 GB card), the Qwen3-VL 32B encoder in NVFP4 (15.7 GB), the int8 video VAE (2.8 GB), the fp32 audio VAE (0.6 GB), and the 2 GB Turbo LoRA unless your model is already distilled. Don't hand-download into the wrong folders - open the ◆ Models card and use its buttons. The URLs are whitelisted, the files land in ComfyUI's own folders, and interrupted downloads resume from a .part file.

Where people get burned

• You're on the Free edition. Covered above. The error text is explicit, so read it rather than reinstalling.
• 16 GB cards. The README's advice is right: W6A8 model, pip install sageattention into ComfyUI's Python, explore in Draft, render the final at 1 MP and upscale after.
• A stale browser cache after updating. The node UI is JS; restart and hard-refresh, or you'll be looking at the old widgets.
• The licence. The MiniMax H3 Community License excludes the EU, UK, South Korea and the US from its applicable territory for the local weights. That's a legal question, not a technical one, and it applies here exactly as it applies to the stock H3 nodes.