Can AI replace fiddly write / record / render continuity workflows? The answer is “maybe, it depends”.

Here’s an experimental piece of continuity for an imaginary TV channel. The design department is on strike and the planning department is on strong medication.

The script and images come from OpenAI, and the voice from Amazon Polly. I integrated them with my data-to-After Effects bulk graphics application. A machine made this, it’s not hand-stitched.

I fed the AI with the names and EPG text of upcoming programmes. Then I asked it for three images and 60 words of VO. The script went to Polly, which sent back an mp3. My application injected all that into an AE project, rendered in EC2, and dropped a proxy back to its dashboard.

So, can AI replace continuity workflows?

The script gets a “could do better” although it’s a passable first stab. Work on the prompt could improve it, perhaps by making it shorter with less room for vamping.

Better voices exist than basic Polly. An integration with something like Veritone could be a big improvement.

The images, not just in this example, were almost always unusable. Better prompting might help, so might a different image generator. Even so there’s a constant risk of the AI hallucinating in a brand-busting object. In another test it gave a cute children’s TV character a stabby knife to hold, because the episode took place in a kitchen.

To use this kind of thing for real, I’d want to ditch the image generation and improve script & voice. Crucially it needs a human in the loop to polish the scripts before handoff to the text-to-speech machine.

On the upside, the cost of this continuity item was pennies. My application could make hundreds more and deliver them automatically every day without ever getting bored.

Download the finished video here.