MiniMax H3 Text-to-Video

The headline feature of MiniMax H3 - and of this node - is that it produces the picture and the soundtrack in a single pass, at 24fps, with native stereo audio. For years, synchronized audio was the one thing the open world couldn't touch: Veo 3's native audio was the capability gap closed models kept. H3 is the model that closed it, and this node is the hosted route to it through the WaveSpeed AI API, so you get it without downloading a 14-billion-parameter model or finding a GPU that fits one. Type a scene, get a clip that sounds like what it shows.

The Audio: line is the trick to steer the soundtrack, and it's not a toggle - it lives inside the prompt:

A drone glides over a rainy city street at dusk, slow push-in.
Audio: rain on a tin roof, distant thunder

The node's own tooltip gives you that exact pattern. Describe the scene and camera in plain text, then add an Audio: line and the model scores the sound to match. Because audio is generated natively, you don't need a separate Foley/sfx step in the workflow - this is the "one node replaces three" case.

How it works

Under the hood it's a thin wrapper over WaveSpeed's hosted H3 endpoint. The node sends your prompt, duration, aspect ratio, and resolution to the API, then polls every 10 seconds until the job finishes (15-minute timeout, generous for 3–15 second clips). What comes back is a URL - the node returns it as a STRING, which is a ComfyUI convention that trips people up: you don't get an IMAGE tensor or a file, you get a link you feed to a preview node.

One quirk worth knowing: the duration snaps to the model's frame grid, so a 5 lands around 5.2s and a 10 won't be exactly 10. That's the API's doing, not a bug.

Inputs that matter

• prompt - required. Scene, action, camera movement, plus the optional Audio: line.
• duration - 3–15s, default 5. Snaps to the frame grid.
• aspect_ratio - seven options (16:9, 9:16, 1:1, 4:3, 3:4, 21:9, 9:21), so it covers portrait/vertical posts as well as widescreen.
• resolution - 480p or 768p. The tooltips give you the price maths: roughly $0.04/s at 480p, $0.10/s at 768p.
• seed - a real generation seed, sent to the API. A fixed seed reproduces the same video and lets ComfyUI serve the cached result without re-billing; -1 (randomize) generates a new one each queue. This is different from the Seedance 2.5 nodes in the same pack, where the seed is cache-only - here it actually shapes output.

The output is video_url (STRING). Wire it into the pack's Preview Anything node to preview or download, or chain it into another node that takes a video URL - the Seedance 2.5 Video Edit and Extend nodes accept it directly.

Installing and configuring

It's part of the ERPK Collection pack. In ComfyUI Manager, search erpk and install ERPK Custom Nodes, then restart. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

The WaveSpeed nodes need a WaveSpeed API key from wavespeed.ai. Put it in Settings > ERPK > API Keys (right-click canvas > ERPK Settings) - stored per-user, never baked into the workflow. No model downloads, no VRAM pressure; the only "heavy" thing is the per-second billing.

Common issues

• Nothing shows up after the run - you're probably piping video_url into a normal preview. Use Preview Anything; it auto-detects video URLs and adds a download button.
• Audio missing or wrong - the soundtrack follows the Audio: line. No line, and the model makes its own call; if you want specific sound, write it down.
• "WaveSpeed" confusion - this is the wavespeed.ai API platform, not the local inference-acceleration library of the same name that speeds up Flux/LTX locally. Different thing entirely; this node needs the API key, not a custom kernel.