Quick answer: the commerce style is image-to-video only, runs on 480P and 768P, and refuses any job without a reference image. It trades flexibility for subject preservation, so the prompt should describe camera, light and motion rather than the product, because the reference already is the product.
Most image-to-video jobs treat the reference frame as a suggestion. Give a model a photo of a bottle and a prompt about a sunlit kitchen, and you will often get a bottle-shaped object in a kitchen. The label moves. The proportions drift. The cap changes colour. For a film still that is fine; for a product shot it is unusable.
H3 has a generation style for exactly this problem. It is called commerce, and it is the strictest of the three styles the generator exposes for this model.
What the commerce style enforces
| Rule | Value |
|---|---|
| Mode | image-to-video only |
| Reference images | at least 1, up to 9 |
| Tiers it runs on | 480P and 768P |
| Duration | up to 15 s |
| Reference audio | up to 3 clips |
| Prompt | required |
| Seed | accepted |
The two that matter: no reference, no job, and no text-to-video. A commerce job always starts from something you supplied. That is a hard constraint rather than a preference, and it is the reason the style behaves differently from a plain image-to-video request with a reference attached.
Why the extra constraints make it more predictable
A general image-to-video request optimises for a good-looking clip. A commerce request optimises for the subject surviving the clip. Everything else in the style — the allowed tiers, the mandatory reference, the duration range — follows from that priority.
In practice it changes three decisions:
- What the prompt should describe. Not the product. The product is in the reference; describing it again in text invites the model to redraw it. Describe the camera, the light and the motion.
- How many references you should send. One clean, well-lit product shot outperforms five near-duplicates. Additional references help when you need different angles of the same object, not when you need the object to look nicer.
- How long the shot can be. Fifteen seconds of a static product is a lot of screen time; the useful range is usually three to six seconds, which keeps the price at 3–12 credits at 480P.
A prompt that works and one that does not
Weak, because it re-specifies the subject:
A sleek matte black skincare bottle with a white cap stands on a marble counter, slowly rotating as morning light moves across it.
Stronger, because it hands the subject to the reference and spends its words on what changes:
Slow dolly right at counter height, camera settles as the light shifts from window light to a warm rim; the product stays framed centre, shallow depth of field, no cuts.
The second prompt contains nothing about the product's appearance. That is the point.
When should you not use it?
You have no flat, clean product image. Commerce needs a reference worth preserving. A phone photo of a bottle on a cluttered desk will keep the clutter, because the style's job is to preserve what it was given.
Your shot needs to move the camera faster than the product allows. Fast moves and heavy parallax are where preservation is most likely to break; if the shot must orbit, test the first three seconds before committing to fifteen.
You are selling a feeling rather than a thing. For a mood piece, a standard image-to-video job with a looser reference is cheaper in attention and no worse in output.
FAQ
Which tiers run the commerce style? 480P and 768P only.
Can commerce generate from a prompt alone? No. It is image-to-video only and requires at least one reference image.
Should the prompt describe the product? No. The reference already carries the product; describing it again invites the model to redraw it.
Where this runs
The commerce style is the third option in the H3 generator's generation-style selector, next to standard and multimode; it only appears on tracks where it can actually run, and the panel restricts it to image-to-video with a reference attached. The tier rates and the reference limits are on https://callirra.com/minimax-h3.
I work on Callirra, so the constraints above are the ones the product enforces rather than conventions I inferred. If a commerce shot drifts on you anyway, the useful report is the prompt, the reference file and the tier — with those three I can reproduce the failure and say whether it is the style, the reference or the prompt.