Generated Video Ads Where the Product Stays Real

Generated video cannot hold an object's geometry, and that single fact decides where we use it. Run the same headboard through a current text-to-video model twice and the slat count changes. Run a car and the wheel arch breathes across the shot. So on this desk the product is never generated. It is photographed, or it comes off your CAD with a matched camera, and generation builds the world around it: environments, weather, transitions, coverage you did not shoot and cannot afford to travel for. How the shots are built: Plates first. A product on a stage, a clean render, or a frame pulled from footage you already own. Generated material is produced as background and mid-ground elements, then composited under and around the product with the camera move solved rather than eyeballed. Where models genuinely hold — smoke, water, foliage, crowd at distance, a city block passing a window, a sky that would take three days of waiting — we use them and it saves you a location. Where they do not, we say so before you have paid for the attempt. Tools turn over every quarter and we have no favourite. What does not change is that a generated frame is treated as an element and never as a delivered shot. Disclosure is now a legal question: Since 2 August 2026 the EU AI Act's transparency obligations have applied to synthetic audio and video, and Meta and TikTok require their AI-content label independently of that. Everything leaving here carries embedded C2PA content credentials, so provenance travels with the file and your team can open any frame and see what touched it. Whether your category carries risk beyond the label is a conversation with your counsel, and we hand them the provenance data to have it with. Hard limits: No synthetic people presented as customers, employees or testimonial subjects. No voice cloning, and no likeness of a real person without written consent you hold and show us. No generated imagery of a product the buyer will later sit on, hold or drive — the gap between the ad and the object is a refund request, and in Turkey a Reklam Kurulu complaint before that. No generated interior sold as a real property. And no retouch request arriving dressed as AI work, such as making the mid-range sofa read as the one two prices up. Not offered: Still image generation. Avatar and spokesperson platforms, at any budget. Generated music or voice. Model fine-tuning on your product library sold as a deliverable — we may train a LoRA to do one specific job inside a project, and it stays inside that project and is deleted with it. The wrong reason to buy this: Price. At this level it is not cheaper than a competent shoot; it is a way to get shots a shoot cannot reasonably give you. If a half-day with a crew would do the job better, we will tell you that during the viability pass and bill you for the pass. Also wrong: any regulated category where nobody has yet asked legal where the line sits, because we will ask before we start and we would rather ask early. Viability pass first: Before a full quote, we take your product, your reference and your placement and produce two or three test frames or a single element, to find out whether this approach earns its place at all. Sometimes the answer is that you should shoot it, and the engagement ends there. If it proceeds: look development and boards, plates, generation and composite, then grade and mix, with two revision rounds counted at the shot level.