The question "should this video be generated or shot" has no answer, because it is asked at the wrong level. A video is not a unit of decision: it is a sequence of shots, and those shots do not carry the same requirements.
The right question is asked shot by shot, and it holds in five words: does this have to be true. A sky, a table top, a kitchen atmosphere do not need to be true to do their job. A hand opening your packaging does.
The decision is taken on the shot list, not on the project
Take an existing brief and look at it as a list. Out of twelve shots, three or four usually carry the whole demonstration, and eight are there to set up, connect or close.
Generating the eight and shooting the four costs less than shooting everything, and produces a more credible video than generating everything. That split is what gets decided, not a camp.
The only case where the question becomes global again is the recurring lead character, because it engages someone's identity. That is a contract subject before it is a production subject, and it is covered in cloning a creator's face and voice. This guide stops at the shooting decision.
What generation handles today
Shots with nothing to prove
Settings, transitions, wide shots, backgrounds, location atmospheres, textures. Everything that sets a tone without proving anything. A sunlit kitchen does not need to exist for a recipe to be credible.
Add animated graphics, packaging elements, and neutral voice overs reading a script without claiming to testify.
What repairs badly afterwards
Generation also produces almost good shots, and that is the real budget trap. A shot whose hands are wrong, whose product label is illegible or whose logo is approximate cannot be corrected: it is redone, or it is cut.
The real cost is therefore not the price of a generation, it is the number of attempts before a usable shot, and that is measured on a test before committing a whole campaign.
What it cannot make verifiable
Three families of shots fail, and they are precisely the ones that sell.
Actually handling the product
Opening, screwing, spreading, pouring, folding. The buyer is looking for physical information there: resistance, texture, scale, how long it takes. A generated shot can show a gesture, it cannot prove one, and the eye catches the gap faster than people think.
It is also the shot where an error shows most: a product whose proportions change between shots ruins the credibility of the whole sequence.
Testimony
A person saying what they thought of a product is only credible if they exist. A generated testimonial is not a weak testimonial, it is something else: an advertising script spoken by an image.
Before and after
The format rests entirely on both images coming from the same person, the same place and the same product. Generating one of the two destroys the demonstration, and generating both makes it worthless.
It is also the format most closely looked at when a claim is challenged, because it asserts a result. A retouched or generated image is doubly expensive there: it weakens the proof and it exposes the message.
Sorting the shots of a brief, concretely
The sort is done in one pass, with a single question per row.
| Type of shot | Does it have to be true | Why |
|---|---|---|
| Setting, mood, transition | no | it sets up, it demonstrates nothing |
| Product alone on a neutral background | yes | the label, the material and the proportions are the information |
| Hand handling the product | yes | the buyer reads resistance and scale in it |
| Real time gesture, duration of an operation | yes | the duration itself is the proof |
| Testimonial to camera | yes | credibility rests on the person existing |
| Before and after | yes | the demonstration rests on both images matching |
| Neutral voice over on product shots | no | it reads a script, it attests nothing |
| Crowd shot, aerial view, distant place | no | it situates, it does not prove |
By contrast with a global arbitration, this reading almost always gives a mixed result, and that is the sign it was done properly. A brief that comes out entirely on one side has rarely been sorted, and it is usually worth a second pass before anyone books a shoot.
What a mixed edit really costs
Mixing generated and shot material is not free, and the cost is not where it is expected. It is not in the generation, it is in the match.
Three constraints come back every time, and all three are settled at the brief:
- Light. A setting generated in late afternoon behind a product shot under fluorescent tubes will not match, and correcting either costs more than having decided beforehand.
- The product's colour, which must stay identical from shot to shot, while generation naturally drifts from one image to the next.
- Framing, because a generated shot knows nothing of the focal length and camera height used on set.
The practical consequence is that the sort has to happen before the shoot and not at the edit. A mixed brief decided upfront asks the creator to shoot on a neutral background or in a known light, which costs nothing extra. The same brief decided afterwards means reshooting, which costs a second day.
It is also why the sort belongs to the brand and not to the supplier. The creator knows how to shoot what is asked; they cannot know which shots you intend to generate if nobody told them.
What has to be said to the viewer
The technical sort does not settle transparency, which arises even when generation only served for the setting.
The European framework on artificial intelligence provides for information duties where content is generated or manipulated, and the spirit of it is simple: someone watching should be able to know what they are watching. In practice, the call to make before release is whether a reasonable viewer could believe that what they see was filmed, when it was not.
A generated setting behind a real product misleads nobody about the product. A fully generated person declaring they used the product does. Between the two, the rule of conduct holds in one line: whatever carries the proof must be true, and whatever is generated must not pass itself off as proof.
The objection you hear is that disclosure would hurt performance. It rarely holds up, and it weighs the wrong risk: a video identified as generated is still a video; a video discovered to be misleading costs trust across the whole brand.
Sources
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence
- CNIL, artificial intelligence and data protection
Checked on 11 September 2026. This guide is not legal advice. Where this guide and the official source disagree, the source prevails.
