MiniMax H3 360 Orbit · Image to Video

The node with no prompt box

Look at the inputs and you'll notice something missing: there's no prompt. MiniMax H3 360 Orbit does exactly one thing - it flies a camera a full 360° around whatever is in your photo - and it doesn't want your opinion about how. You give it an image, a resolution, a duration and a seed, and you get a loop back.

The loop is the clever part. Your photo is used as both the first and the last frame, so the clip ends where it started and repeats seamlessly: gif, product turntable, character showcase, background plate. That's the tooltip's claim and it's the whole reason this node exists separately from the general H3 video model. Nobody wants a 360 orbit that snaps back to a different image at the end.

Where you'd use it: a product photo that needs to feel dimensional, a character sheet that needs to read as a 3D object, a portrait with a spinning-background social-media look. Where you wouldn't: anything that needs acting, dialogue, or a shot that goes somewhere. There is no room in this node for direction, which is both its limitation and the reason it's two minutes of work instead of an afternoon.

How it runs

Upload the image (PNG, or JPEG if it's over the 10 MiB ceiling), quote, submit, poll, done. The clip comes back as a standard VIDEO - Save Video it. H3 is MiniMax's 33B omni-modal model, and the open weights made a lot of noise in August 2026 for generating video and stereo audio in one network. Here you're calling the hosted endpoint, so none of the licensing geography around the open weights applies: no download, no community-licence territory question, just credits.

Since the model is orbiting your exact image rather than re-imagining it, orbit quality tracks photo quality. A clean subject on a separable background beats a busy snapshot where the camera can't tell what to circle.

Inputs

• image_url - the photo to orbit, and the loop's anchor at both ends.
• resolution - 480p default, then 540p, 768p, 1080p. Price climbs steeply per tier: roughly $0.05/s at 480p up to $0.20/s at 1080p.
• duration_seconds - 3 to 6 seconds only, default 3. The tooltip is unusually direct: three seconds gives the cleanest frozen orbit. Longer doesn't buy smoothness, it just gives the model more time to invent wobble.
• seed - fixed by default. Two identical queues, one bill.

Outputs: video and task id.

Because durations are short, this is a cheap node to experiment with. Three seconds at 480p is around fifteen cents; three at 1080p is about sixty. Perfect the resolution last.

Install

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

Manager users can search PopNow and install from there; the CLI has comfy node install popnow-comfyui. No dependencies, no models, no ffmpeg - the pack uses only what ComfyUI ships, and the point of a cloud node is that the 33B stays on someone else's rack.

Then a key: popnow.ai/api/console/keys, pasted into Settings > PopNow, or POPNOW_API_KEY set in ComfyUI's environment. Restart once afterwards so the node list is rebuilt from your account's catalog - that fetch is also what makes models appear when PopNow adds them, so a missing node is nearly always a missing restart.

Things that catch people

Expecting prompt control. There isn't any. If you want the camera to arc, pause, push in or change direction, this is the wrong node - the H3 image-to-video node takes a prompt.

A visible seam. If your photo has strong directional lighting, the back of the orbit is being invented, and it shows. Front-lit, evenly lit subjects loop better.

Duration creep. Six seconds isn't twice as good as three. Start at three, and only go longer if the orbit is genuinely too fast to read.

A loop that judders at the join. That's usually the source image rather than the seed; try a tighter crop so the model has less to invent outside the frame.