MiniMax H3 Max · Text to Video

Here's the thing nobody mentions in the teaser clips: this node's aspect ratio defaults to 1:1. Square. You queue your first epic widescreen shot, wait, and get a square video with a soundtrack. It's a small trap, but it's the first one everybody hits, so set aspect_ratio before you fall in love with a prompt.

What it is

The text-to-video node for MiniMax H3 Max in the PopNow pack - cloud inference, no local weights, from about 7.5 credits ($0.075) per second of output. H3 Max is the speed-oriented variant of MiniMax's omni-modal video model: native audio generated alongside the frames, 5 to 15 second clips, 480p or 768p. It's not the quality ceiling of the family (base H3 does 2K), it's the one you use when you want a usable take in a couple of minutes.

The reason to route through a cloud node at all is worth being blunt about. H3's weights are open, but the licence excludes the US, EU, UK and South Korea from its Applicable Territory, and the full model is a 33B thing that most people can't run anyway. The hosted endpoint has neither problem.

How it works

Four widgets go out as a request: prompt, aspect ratio, duration, resolution - plus the optional prompt-expansion switch. PopNow quotes it, holds at most the quote, renders, and the node pulls the finished MP4 back and wraps it as a ComfyUI VIDEO. Progress, queue position and the final charge are printed on the node itself, which is a nicer experience than watching a console scroll.

Because audio comes from the same pass as the picture, the model responds to sound words. "Distant traffic, a busker's trumpet leaking through the window" gives you something to work with; on a silent generator those are wasted tokens.

The inputs

• prompt - subject, action, camera, mood. Long-form works: the field takes thousands of characters, and this model would rather read a paragraph than a list of tags.
• aspect_ratio - defaults to 1:1. 16:9, 9:16, 4:3, 3:4 and 21:9 are the other choices. Set it first, complain later.
• duration_seconds - 5 to 15, default 8. The default is longer than the base H3 node's, so check the price badge before queueing a batch.
• resolution - 480p or 768p, default 768p.
• enable_prompt_expansion - off by default; turning it on adds a flat per-request fee. Useful for one-line prompts, annoying when you've already written the shot precisely.

video goes to Save Video, task id to your notes. No seed widget on this node, so reproducibility by number isn't on the menu - a design decision that makes sense for a service where every run is billed.

Drafting prompts that survive the expansion switch

The node's tooltip asks for "the subject, action, camera, and mood of the shot," which is a decent checklist. In practice, the prompts that come back usable do four things: name one subject, give it one action with a beginning and an end, state the camera (locked off, slow push, handheld follow), and establish boundaries - where the shot starts and where it stops. A clip that tries to do three actions in eight seconds reads as mush regardless of resolution.

Adding sound to that list is where H3 pulls ahead. One line of ambience plus one specific sound event is plenty; describing a whole soundscape tends to swallow the visuals.

Install

Manager → search PopNow → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart, then paste a key from popnow.ai/api/console/keys into Settings > PopNow and restart again - the node list is rebuilt from your account's catalog at startup, so a model that seems missing usually just needs that second restart. Nothing else gets installed; the pack's dependency list is empty, and it needs Python 3.10+ with ComfyUI's V3 node API.

Troubleshooting

PopNow 40201 is credits. PopNow content_rejected is the model refusing the prompt - failed runs are refunded, so it costs time, not money. PopNow 400 naming a field means an unsupported value; the four combos on this node are narrow and the model page linked in the node description lists exactly what it takes. Errors ending in request_id ... are the ones to forward to support, because that id pins the exact request.