The short answer
Use generated video where nothing has to be exactly accurate, and use a studio where something does. That single line resolves most of the argument. Atmosphere, texture, concept films and social variations are now genuinely fast and cheap to generate. Your product, your interface, your packaging, your typography and your logo are not things a generator reproduces reliably, and a customer comparing the film with the box notices immediately.
What changed in 2026
Text-to-video crossed from demo to production tool over the last eighteen months. Output at short durations is frequently good enough for social placements, the tools are sold through ordinary subscription and credit plans rather than enterprise deals, and 2026 was the year national advertising slots carried work made this way. Small teams now produce material that previously required a crew.
What did not change is the failure mode. Generators are excellent at plausible and unreliable at specific. Physics wobbles, objects change between shots, text renders as approximate lettering, and a product's proportions drift. None of that matters for a five-second abstract background. All of it matters for a hero film about a thing you sell.
Where each approach wins
| Job | Generated video | Studio production |
|---|---|---|
| Concept and pre-visualisation | Strong. Minutes per option. | Slow and expensive for exploration. |
| Product accuracy | Weak. Approximates your object. | Strong, especially from CAD or a built 3D model. |
| Brand assets, logo, type | Unreliable. Letterforms distort. | Exact by construction. |
| Atmosphere and texture | Strong and cheap. | Possible but rarely worth the budget. |
| Consistency across a campaign | Drifts between generations. | Controlled by the project file. |
| Revisions to a specific frame | Re-roll and hope. | Change the layer and re-render. |
| Rights and clearance | Terms vary, protection uncertain. | Contractual and assignable. |
The revision row is the one that decides most real projects. A client note like "same shot, move the label two millimetres, warmer light" is a five-minute change in a studio file and an unbounded gamble in a prompt.
The accuracy problem, concretely
We build 3D product films, so the comparison is not hypothetical. When a beverage brand needs a pour, the geometry of the bottle, the label's exact artwork, the liquid's colour and the way light moves through it all have to match the physical product. A generator makes a beautiful bottle that is not yours. A model built from the real dimensions makes your bottle, and then makes it again next quarter for a different flavour without redesigning anything.
That is also why generated video does not solve the problem we described in why 3D renders look fake. Realism is a function of light, materials and restraint, and a generator's version of realism is an average of its training data rather than a description of your object. It is convincing until it is compared.
For brands weighing renders against photography, the calculus in 3D visualisation versus product photography is unchanged by AI video: the value is in a reusable asset that can be re-shot without a studio day, and a generated clip is not reusable in that sense.
How studios actually use these tools
Quietly, and early in the process. The honest account of a 2026 workflow looks like this:
- Mood and direction. Generate a dozen treatments in an afternoon to agree a direction before anything is modelled. This replaces static mood boards, not the film.
- Pre-visualisation. Rough the camera move and pacing to test a cut before committing render hours.
- Backgrounds and plates. Abstract environments, skies, textures and light studies where nothing specific must be true.
- Variation. Once a cut is approved and built properly, generate supporting fragments for placements that need volume rather than precision.
The hero asset is still built. What has collapsed is the cost of deciding what to build, which used to be the slowest part of a project and the one clients found hardest to pay for.
Rights, disclosure and the things that bite later
Three practical cautions, all of which arrive after the excitement.
- Licensing. Check commercial-use terms for the specific plan you are on, and keep evidence of what was generated where. Terms have changed repeatedly.
- Copyright. Purely machine-generated output may not be protectable in several jurisdictions, which matters if the asset is meant to be defended. Human authorship in the chain is not a formality.
- Disclosure. EU rules now require marking of synthetic output and disclosure of AI-generated depictions of real people, places and events, and platforms add their own labels. We covered the rules in the EU AI Act's transparency obligations, and the customer-trust side in whether customers care about AI disclosure.
There is a brand argument underneath the legal one. In a feed where everything is generated, the recognisable signature of real craft is becoming the differentiator — the same dynamic we described in why every website looks the same.
How to decide on a given project
Ask one question: does anything in this film have to be exactly right? If the answer is a product, an interface, a logo, a price, a legal claim or a real person, build it. If the answer is a feeling, generate it, and generate ten versions while you are there.
And whatever the source, the first three seconds still carry the work, and the file still has to be light enough to play. Those constraints are older than the tools, and we wrote them up in the first three seconds and choosing an animation format. If you have a film to make this quarter and want an honest answer about which half of it should be generated, show us the brief.