MiniMax H3 LoRA · Text to Video

Prompt to clip, with your adapters and its audio

MiniMax H3's text-to-video, with the LoRA stack attached. Three adapters, a seed you control, four resolution tiers from 480p to 1080p, and up to 15 seconds of output with stereo sound generated in the same pass as the picture. That last part is the reason H3 got the reception it did: it treats audio as part of the modality rather than an afterthought, so a clip arrives with a voice, a room tone or a beat already attached.

The 480p tier is the part I'd underline. At around $0.05 a second, a 5-second draft is a quarter. H3's cheaper tiers exist precisely so you can iterate - explore the prompt and the LoRA scale at 480p, then take the winner to 768p or 1080p once the shot is worth the money. The base H3 text-to-video node can't do that; it only offers 768p and 2K, which makes every experiment cost real money.

Compared to its base sibling, this node also adds the seed. That means identical re-queues reuse the cached result instead of billing you again, and it means you can nudge one word while everything else stays put - actual iteration instead of re-rolling.

Mechanism

Prompt, widgets and any LoRA references go out; the job is quoted first, submitted with an idempotency key, and polled until the file lands. LoRA files were resolved upstream by PopNow LoRA: local .safetensors files from ComfyUI/models/loras are uploaded once and cached for twelve hours, Hugging Face and Civitai references are imported server-side, and direct links get passed through. The output is a VIDEO - into Save Video, or anywhere downstream that eats video.

Prompt structure matters, and the tooltip is specific about it: describe the scene, the action and the camera; put the soundtrack on its own Audio: line; put LoRA trigger words first. H3 was trained to think of text, image, video and audio as one context, so treating sound as a separate prompt field is using it the way it wants.

Inputs

• prompt - scene, action, camera, then Audio:.
• loras - up to three H3 video adapters, each with a scale from -1 to 4.
• aspect_ratio - 16:9 default, and the list runs wider than the image nodes: 9:16, 1:1, 4:3, 3:4, 21:9, 9:21.
• resolution - 480p default, then 540p, 768p, 1080p.
• duration_seconds - 3 to 15, whole seconds.
• seed - fixed out of the box.

Outputs: video and task id.

Install

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

Restart. Or Manager > search PopNow > Install; or comfy node install popnow-comfyui. The pack declares no dependencies at all - no torch, no CUDA, no model downloads - because the model runs on PopNow's servers and the node is just a well-behaved API client. That's the whole reason a 33B video model can live in your custom_nodes folder.

Then a key: popnow.ai/api/console/keys → Settings > PopNow, or POPNOW_API_KEY in the environment (which wins and keeps the key out of anything you might share - the pack never writes the key into a workflow or image metadata). Restart once more so the node list comes from your account's catalog.

Troubleshooting without hand-waving

The LoRA doesn't show up in the output. Check the family first: H3 video adapters only. Then check the trigger phrase is in the prompt, early. Then the scale - 1.0 is a default, not a recommendation, and stacked adapters usually want less each than they'd take alone.

Sound is wrong or absent. Say what you want. H3 will invent a soundtrack if you don't ask, and invented audio is often generic. The Audio: line is in the tooltip for a reason.

Import failures before generation. Gated Civitai or Hugging Face files need CIVITAI_API_KEY or HF_TOKEN in ComfyUI's environment. Tokens are environment-only by design; a widget value gets saved into every workflow you share.

Wallet errors. PopNow 40201 means not enough credits - buy some or drop to 480p. PopNow 40004 means the settings combination isn't servable and the message names the field. content_rejected means the model declined and the run was refunded. And a cancelled run isn't refunded: the job finishes on PopNow regardless, and you're charged if it succeeds.

Set a ceiling before you scale up. Settings > PopNow > Max credits per run refuses any quote above your limit before credits are held. It's the cheapest insurance in the pack, and this is the pack where per-second pricing makes it worth having.