The request failed the first time because I had not attached the product.
I had written a beautiful prompt about a glass bottle on a wet slate surface and pressed generate, and the commerce path refused it — text-only is not allowed there. My instinct was that this was a limitation. Two hours later it was the reason the batch worked.
The short version. The commerce path renders 1–15 s at 480P or 768P, 16:9 or 9:16, and requires at least one product reference image. Text-only requests are rejected.
What does the commerce path allow?
480P or 768P, nothing higher. 16:9 or 9:16. One to fifteen seconds. At least one reference image, always. No scene presets.
A 6-second 768P clip costs about twelve cents. I rendered nineteen of them across seven products before lunch, which is the kind of sentence I could not write about any other path.
The four shots that pay for themselves
The turn. Four to six seconds, product rotating a quarter turn, light sweeping across the label. One reference, one instruction: label stays readable, soft key from the left. This is the shot that replaces a still on a product page, and it is the one that fails the most visibly if the label blurs.
The pick-up. Six to eight seconds. A hand enters, lifts, sets down. The hand is what goes wrong — frame it small, say where it enters from, and do not ask for a wristwatch you did not reference.
The before and after. Eight to ten seconds, one continuous move across two states. I now build this as two takes joined later rather than one prompt that tries to morph; the morph always looks like a dissolve of two products.
The in-use. Ten to fifteen seconds, hands and context, no face required. The longest and the most convincing, because motion covers for detail.
Why the mandatory reference helps
Without one, the model invents the product. An invented product is wrong in ways that are hard to name and impossible to fix by rewording — a label with plausible but incorrect letterforms, a cap with the wrong ridge count.
With one, the failure mode narrows to "the motion is unconvincing", which is a normal production problem with normal fixes: shorten the shot, simplify the move, change the light.
So the workflow is: one clean still per SKU, even light, label square to camera, saved as the reference. Every shot for that SKU comes from it. Extra references are free at these tiers — the per-image surcharge only exists on 2K.
When is the commerce path the wrong choice?
1080P delivery — the path stops at 768P, so a hero spot belongs on another tier. Text-only concepts — there is no version of this that runs without a picture. And complex choreography: two actors and a product is more than a 768P selling clip needs, and it is where the scene presets belong instead.
What I would tell a small brand
Shoot the still properly. Not a phone snap against a window — a clean, evenly lit frame with the label readable and the silhouette uncluttered. The whole batch inherits that picture's problems, and no prompt is going to give the product back its edges.
The commerce generation style and its limits are on the H3 page.
I work on Callirra, so these limits are the path's own rules rather than my guess at them. If you have a packshot where the label holds better than mine does, the useful thing to publish is the reference you used.