MiniMax H3 Text to Video

This is the headline node of the pack, and the honest pitch is in the difference between "MiniMax H3" and "MiniMax H3 API." Running the actual H3 weights locally means pulling down a 33B omni-modal model - about 42GB - and the license that came with it is the one that geofences out the US, EU, UK, and South Korea, even for the outputs. This node does none of that. It calls MiniMax's hosted H3 API, which is globally available, needs no GPU on your side, and drops the finished clip back into your workflow as a normal VIDEO output. You trade local control for a working prompt-to-video pipeline that runs on hardware you don't own.

How it works

The node is an async API client dressed up as a sampler. When you hit Queue it validates that the prompt isn't empty, builds a JSON payload with your prompt, resolution, duration, and ratio, and POSTs it to <base_url>/v2/video_generation. The API replies with a task ID, and the node then polls /v2/query/video_generation/{task_id} every 10 seconds until the video is ready (hard cap at 30 minutes), then downloads the result and hands you a VIDEO output. Between Queue and result you're looking at real wait time - this is a cloud render, not a 30-step local denoise. A 5-second clip usually comes back in a couple of minutes; long 2K clips can take a while.

What you're buying is worth saying plainly: H3 is the model that opened the first serious open-weights answer to Veo's native-audio gap. It treats text, image, video, and audio as one input context and generates clips with synced stereo sound - up to 15 seconds at 2K/24fps. Text-to-video is the simplest entry to that.

The inputs that matter

• prompt - your text description. Required, and the only input with real creative leverage.
• ratio - aspect ratio, default 16:9. For text-to-video this is not optional in the adaptive sense: the node's tooltip says it plainly - ratio cannot be adaptive, there's no image to infer it from. Pick from 16:9, 4:3, 1:1, 3:4, 9:16, 21:9.
• resolution - 768P or 2K. 768P is cheaper per second; 2K is where H3 shines. Default is 768P.
• duration - 4–15 seconds (the slider goes 4–15; the README's "5–15" is slightly stale).
• watermark - off by default. Flipping it on requests an AIGC watermark.

There's also a seed input, and here's a gotcha worth knowing: the source comments say it plainly - seed is accepted by the node but not sent, because the v2 API has no seed field. It's a control that does nothing. Same seed, different seed, the API treats them identically. Don't waste time "tuning" it.

The output is a single VIDEO - in recent ComfyUI, a proper video type you can wire into a video player, a SaveVideo, or frame-extraction nodes, not a bunch of latent tensors.

Install

ComfyUI Manager → Install via Git URL → https://github.com/meta-sota/ComfyUI-MiniMaxH3-API.git, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/meta-sota/ComfyUI-MiniMaxH3-API.git
cd ComfyUI-MiniMaxH3-API
pip install -r requirements.txt

Only pip dependency is aiohttp. The real requirement is a modern ComfyUI (author tests 0.30.0) because the pack is built on the native V3 node API. No model files, no VRAM - the Key node in the workflow holds your credentials, and every generation node needs one connected.

Common issues

• "MiniMax H3 token is required." - no Key node wired in, or its token is blank.
• Long, seemingly stuck runs - it's polling. The node logs task status and will error cleanly on failed/cancelled states or after the 30-minute cap. If the queue genuinely errors, the error text comes straight from the API, so read it.
• Empty prompt - rejected before anything is sent, so if nothing happens, check the prompt field.