An AI campaign film: the real problem is character consistency

First attempts at filmmaking with generative models are always encouraging: you write a prompt and a cinematic frame appears in seconds. The problem starts at the second frame. Ask for the same character again and facial proportions, costume detail and colour temperature drift; placed side by side the two frames look like two different people.

So the work is built around character production before video production. Reference images of the character are fixed first — pose, costume, props, light — and every scene is derived from that reference. Motion is generated from those stills rather than straight from text (img2vid). Consistency is then tied to a file you hold, not to the model's memory.

In a ten-part campaign this approach has a second benefit: the production plan. If each scene is cut as an independent 30-second clip, regenerating one scene does not disturb the others. In a single long film an error in the middle means regenerating the whole chain.

Client approval moves into the queue as well: scene synopses and a frame-by-frame storyboard are shared before production. In generative work a 'reshoot' looks cheap but it is not free; approval still belongs before production.

A closing note that is legal rather than technical: if the brand has its own mark, colours and patterns, production must be bound to them. However good the model's aesthetic, a character that ignores the brand guide represents the model, not the campaign.