MiniMax H3 · Text to Video

Type a scene, get a clip with sound already in it. That's the whole pitch, and it's a better pitch than it sounds: most local video pipelines generate silent footage and then need a second model, a foley pass, and a timeline to feel finished. MiniMax H3 makes the audio in the same generation as the picture, and this node is the front door to it.

What it is

PopNow_minimax_h3_text_to_video runs MiniMax H3's text-to-video endpoint on PopNow's servers and drops the result into your graph as a VIDEO. Nothing is installed beyond the node pack - no checkpoint, no VAE, no 42 GB download, and no GPU requirement at all. It's part of the PopNow pack, which wraps roughly 248 cloud models this way; the H3 nodes are the MiniMax corner of PopNow > Video > MiniMax.

H3 itself is worth one paragraph of context. It's MiniMax's 33B omni-modal model, published as open weights in August 2026 with a licence that excludes the US, EU, UK and South Korea from its "Applicable Territory" - so for a lot of readers, running the weights locally isn't even licensed, never mind practical. The hosted API has no such carve-out. That's a real, if unglamorous, reason cloud nodes like this one exist.

How it works

The node builds a request from your four widgets and nothing else: prompt, aspect ratio, resolution, duration. PopNow quotes the run, holds the quoted credits, renders, and the node downloads the finished file and wraps it as a ComfyUI VIDEO. You'll see the estimate, the queue, the generation time and the final charge printed on the node while it works.

Because audio is generated jointly rather than bolted on, prompts that describe sound behave differently from what you're used to. "Rain on a tin roof, a door creaks open" gets you rain and a creak. On a silent model those words are just atmosphere keywords.

The four inputs

• prompt - picture, action, camera, sound, and the shot's boundaries. Up to 7000 characters, and the long limit is there for a reason: this model rewards a written paragraph, not a tag soup.
• aspect_ratio - 16:9 default, with 9:16, 1:1, 4:3, 3:4 and 21:9 alongside. For anything headed to Reels or Shorts, change this before you generate, not after.
• resolution - 768p or 2k, default 768p. Priced per second of output with the 2K tier costing more; the node badge shows the live rate for whatever you've picked.
• duration_seconds - 5 to 15 in whole seconds.

Outputs: video and task id. Wire video into Save Video. There's no seed widget on this node, so if you get a take you like, save it - you can't dial the same one back in by number. (For seed control and LoRA slots, the H3 LoRA and Singularity LoRA nodes in the same category have them.)

Install

Manager → search PopNow → Install, or:

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

Then restart, add your key at popnow.ai/api/console/keys via Settings > PopNow - or set POPNOW_API_KEY if this is a headless box - and restart once more. Dependencies are nil (pyproject.toml lists none); the pack needs Python 3.10+ and ComfyUI's V3 node API, both of which you already have.

If you used the old single-file popnow_api.py node from popnow.ai's docs, delete it. The pack's PopNow Generate node replaces it, and leaving both installed is a confusing way to spend an afternoon.

Getting burned

PopNow 40201 is an empty wallet - lower the resolution or the length, or top up. PopNow content_rejected means the model declined; failed runs are refunded, so it costs you time rather than credits. The error that ends in a request_id is the one to send to support, because it identifies the exact request.

Two quieter issues. First, this node has no negative prompt, so "no text, no watermark" has to be phrased positively inside the prompt. Second, every run is billed: ComfyUI's queue button will happily charge you again for the same graph. The seeds in this pack default to fixed precisely so a double-queue reuses the first result, but with no seed field here, treat each queue as a purchase.