The H3 mode that refuses to run without a product photo

#video #minimax #ecomemerce #generation

The first time I ran a packshot through MiniMax H3, the job failed before it started.

Not a bad render — a rejection. The commerce path refuses a request that has no reference image attached. No image, no job, nothing charged. My instinct was that the mode was broken. It was not: the hard rule is the feature.

Most video models treat a reference image as an optional hint. This path treats a missing reference as an invalid request, which is exactly what product video needs and exactly what makes the path useless for anything else.

I work on Callirra, where this path is selectable at 480P and 768P. It is the most opinionated mode on the model, and understanding the opinion saves render credits.

What the mode is and is not

	Commerce path	General image-to-video
Reference image	required	optional
Tiers	480P, 768P	480P through 2K
Text to video	no	yes at most tiers
What it optimises	keeping the subject	following the prompt
Reference audio	not part of this path	up to 3 clips

The rule that follows from the middle column: this is a subject-preservation machine. A product shot has one non-negotiable requirement — the object in the output has to be the object in the photograph. Camera, light, motion, environment are all negotiable; the bottle is not.

The call

curl https://api.callirra.com/v1/videos \
  -H "Authorization: Bearer $CALLIRRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3",
    "prompt": "slow 15-degree orbit, soft key from camera left, hard rim behind, condensation runs, seamless warm backdrop",
    "resolution": "768P",
    "duration_seconds": 8,
    "aspect_ratio": "9:16",
    "frame_images": ["https://your-cdn.example/packshot.jpg"]
  }'

Two things worth noticing. The aspect ratio is 9:16 because the deliverable is a vertical ad slot, and the prompt describes how the object is seen, not what the object is.

Why the prompt changes shape

When the reference carries the subject, describing the subject fights the reference. Upload a perfect packshot and then spend four hundred words describing the bottle, and you have given the model two competing descriptions of the same thing.

Describe instead:

Camera — "slow 15-degree orbit, 35mm equivalent, slight parallax against the background".
Light — "single soft key from camera left, hard rim from behind, deep shadow on the right".
Motion — "condensation runs down the glass; the liquid inside stays still".
Environment — "seamless warm backdrop at f/2.8, no props".

The subject is already decided. Everything you write should be about how it is seen.

Where it goes wrong

No reference image. Rejected. That is the mode working.

A hard cut-out on white. Product shots depend on believable edges. A cut-out with a pre-baked shadow tends to keep that shadow through the orbit, and the object looks pasted onto the new scene.

Asking for a scene change and a subject change. Swapping environment while keeping the object is the mode's job. Swapping the object is not.

Expecting audio. No reference audio here, and no dialogue register.

A workflow that stays cheap

Draft at 480P: one credit per output second is low enough to test five camera moves on the same packshot before committing. Once the move is right, re-render the selected take at 768P. A ten-second master at 768P is twenty credits — twenty cents at the published rate.

Keep the winning reference and the winning prompt together. The next product in the range reuses the camera and light blocks verbatim; only the image changes.

When not to use it

If the deliverable needs 1080P or a larger raster, this is the wrong mode — use the general image-to-video path with references attached. If the shot needs dialogue, no. If the object has to change mid-shot, no.

FAQ

Does it work with text only? No. Without a reference image the job is refused.

Is 1080P available? Not on this path; it runs at 480P and 768P.

Will the product be pixel-identical? No path guarantees that. This one is tuned for preservation, and you should still review edges, labels and any text on the product.

Why draft at 480P instead of just rendering the final? Because five camera moves at 480P cost less than one wrong take at the final tier.

The commerce path is selectable in Callirra's H3 generator. If you have a packshot this mode handles badly, the reference and the prompt are the two things worth sharing.