The scene is generated. The product never is.

Why fully generative fails here

Marketing needs a constant stream of short product video. Generative video looks like the obvious answer, and for a brand this size it is the wrong one — the product itself gets re-invented. Shape drifts, proportions drift, the logo comes back almost right. You cannot ship creative in which the product is an approximation of itself.

And the hard part is the back of pack: multilingual directions, the INCI ingredient list, the barcode, regulatory marks. Generative models do not reproduce that — they produce something that looks like it from two metres away.

Hybrid, not generative

Our templates build the shot around a product-grounded 3D asset. The product is deterministic — a model of the real packaging, so it is right by construction. The scene around it is generated. That split is the whole idea. I wrote up the general architecture in hybrid product visualization.

Look at the back of the bottle

This is a frame from the render, not a photograph. Directions for use in five languages, the full INCI list, the batch code — every line is the real text from the real pack, because it came from the 3D asset rather than a model's idea of a label. This is the part no generative pipeline gets right, and the part a regulated category cannot get wrong.

The economics

Once the 3D asset exists, a new variant is a render, not a gamble. A fully generative clip runs about $1–6 per 8 seconds before retries, and the output is probabilistic. Asset-based rendering comes in under $1 for 15–20 seconds, accepted on the first render and identical every time — one to two orders of magnitude cheaper per accepted variant, and faster and more predictable. The cost model is set out in the SaudiCIS 2026 paper on hybrid product visualization.