Generative AI

Generative AI describes models that produce new content instead of merely recognising or sorting what exists. For video that means a shot appears from a description or an image where there was none. The models are trained on large amounts of data and do not necessarily return the same result twice.

What that means for production

Three properties shape the work. First, the result is not predictable, so work runs in loops with a lot of discarding. Second, the model has no memory between two runs, which makes character consistency the actual job. Third, what did not appear in training cannot be rendered accurately, including your specific product with its label.

From that follows the key number: for one usable shot there are at least five discarded. That waste is not a flaw in the method, it is the method, and it is why selection and direction make the difference rather than the prompt.

Which models we use

Seedance, Veo, Kling and Google Flow for moving images, Seedream, Nano Banana, GPT Image and Seed 1.0 for stills, Topaz for upscaling. Which tool a project gets is decided by the subject rather than by habit, and we test new models continuously without putting them into production for that reason alone.

For confidential products we work through access routes where your material is not used to train the models. That costs us more and still does not appear separately on your invoice. Which tools a project used is disclosed on request.

Frequently asked questions

What is generative AI?

Generative AI describes models that produce new content from an input, such as images, video, audio or text. They differ from recognition systems in that the result did not exist before and can vary for the same input.

Do the models learn from our material?

That depends on the access. With public services it is possible depending on the contract, through enterprise interfaces it is generally excluded. For confidential products we therefore use the route where input does not feed training, and record that in our terms.

Will generative AI soon do everything?

The models are improving quickly, especially on movement and consistency. What they do not replace is deciding what should be said and which of a hundred results is the right one. That work does not disappear, it becomes more important.

Read on: All terms in the glossary • What we produce with

Which model suits your subject?

We answer that in the first call, not after the invoice.

Book a first call