MiniMax H3 · Image to Video

The base H3 video node, and its two surprises

MiniMax H3 is the 33B omni-modal model that made everyone sit up in August 2026: text, image, video and audio in one context, native stereo sound, clips up to 15 seconds at 2K. This node is its image-to-video entry in ComfyUI - you give it an opening frame and a description of what happens next, and you get a clip with sound.

Two things about this particular node differ from its siblings, and both matter more than they look.

There's no seed widget. The plain H3 image-to-video doesn't expose one. So every queue is a genuinely new render, and every queue is billed - no caching a good result by re-running identical settings, and no "same seed, tweak one word" iteration. Get the prompt right before you press the button, or accept that you're paying per roll.

There's no loras socket either. For adapters you want the MiniMax H3 LoRA image-to-video node, which also adds a seed, more resolution tiers and a wider duration range. This node is the base model, cleanly.

What you do get is an optional closing frame: last_image_url. Connect one and the clip has to resolve on it, which is how you bookend a shot - start here, end there, let the model negotiate the middle.

How it works

Your opening frame is encoded (PNG, or JPEG above the 10 MiB ceiling) and uploaded, the job is quoted and submitted with an idempotency key so retries don't duplicate, and the node polls until the MP4 arrives as a VIDEO value. Save Video it and watch it in a real player; ComfyUI's preview is not where you judge a 2K clip.

Prompting advice that's worth more than it costs: the tooltip asks for the action, the camera and the sound. H3's whole thesis is that audio and picture are generated together, so "she looks up, camera drifts in, a door closes somewhere behind her" is using the model as designed. Leave the sound out and you get whatever the model invents.

Inputs

• image_url - the opening frame. The tooltip calls it an "Authorised opening frame", i.e. it's your picture to upload.
• last_image_url (optional) - the closing frame. The one real control on the node.
• prompt - action, camera, sound.
• resolution - 768p default, or 2k. Two tiers only. Roughly $0.037/s at 768p and $0.058/s at 2k.
• duration_seconds - 4 to 15, whole seconds, default 5.

Outputs: video and task id.

Install

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

Manager > search PopNow > Install also works, and so does comfy node install popnow-comfyui. There's nothing else - no model to fetch, no dependency to satisfy. A 33B model as a node with a two-line install is the entire value proposition.

Then a key from popnow.ai/api/console/keys, pasted into Settings > PopNow or exported as POPNOW_API_KEY. Restart once after that so the node list gets rebuilt from your account's catalog; that's also the step people forget after an update, and then wonder where a model went.

Watch out for

No seed means no cheap retries. Every run is a fresh charge. Set Settings > PopNow > Max credits per run to a limit you're comfortable losing, and draft short - 4 seconds at 768p is about fifteen cents, 15 seconds at 2k is closer to ninety.

Duration limits are quirky. This node runs 4 to 15 seconds, where the LoRA sibling starts at 3. If you specifically need a 3-second clip, you're on the wrong node.

Bookending isn't a guarantee. last_image_url aims the clip at your closing frame; it doesn't promise frame-exact delivery. Expect a soft landing rather than a snap.

Cancelling doesn't refund. ComfyUI's cancel stops your wait. The job keeps running on PopNow and is charged if it succeeds. Failed or refused runs (generation_failed, content_rejected) are the ones that come back refunded.