When the AI ​​creates the movie trailer

Share your love

It’s the old cliché: If there’s one thing artificial intelligence isn’t, it’s creative. For example, computers cannot come up with stories that are worth reading, they say. But after the triumphant advance of machine learning (ML) and text generators based on large language models based on it, it is becoming more and more evident that AI systems can also become authors – even if this only succeeds because they now perfectly imitate human writers and texts can thus “predict”.

A team from the University of Edinburgh has now trained an AI system to create movie trailers. Such systems already exist, but their output is still in need of improvement. That is why the Scots are using a new approach that combines old models with improved data sources. The end result was so good that human evaluators at the crowdworking service Amazon Turk preferred those variants from the computer whose models had been trained completely unsupervised (“Unsupervised ML”), i.e. came from the new algorithm.

More from MIT Technology Review

More from MIT Technology Review

More from MIT Technology Review

More from MIT Technology Review

The team consisting of Pinelopi Papalampidi, Frank Keller and Mirella Lapata from the Institute for Language, Cognition and Computation at the Computer Science Faculty of the University of Edinburgh shows in his preprint paper their neural network based approach. To analyze an existing film, both its script and the video are used; A graph model is then created from the film. The system can identify the narrative structure and mood of individual sequences as well as important segments (“turning points”) in the script according to film theory – such as setbacks for the main actor, the “call to action”, highlights of the story or the grand finale.

By identifying the mood of areas of the film, the trailer from the computer can conform to certain narrative concepts of how trailers should generally be constructed. This includes, for example, that the film previews often begin with a scene of medium intensity. However, so far the system has only been able to identify basic moods, but not the characters’ finer-grained feelings – such as fear, joy or sadness. A total of 41 trailers were developed by the AI ​​system.

The Scottish researchers’ system is not yet so good that it could replace human creatives – but it seems to beat previous systems that have been used to generate trailers, including CCANet for Trailers and the supervised, trained variant of the Graphtrailer model. “We model films as graphs in which the nodes are individual settings and the edge areas represent the semantic relationships between them,” said the researchers. The system could one day help to at least accelerate trailer production.


(bsc)

Article Source

Read Also   betterCode () API: Presentations wanted for the conference on web API development
Share your love