MOVIOLIA uses AI to boost video editing

Artificial intelligence is beginning to shake up a wide range of professions and sectors, including the most unexpected, such as... video editing. This is confirmed by the Moviolia project, on which IMT Atlantique's BRAIn (1) team is working in partnership with production company Tydeo and Ektacom, a specialist in digital storage and video streaming. 
The idea behind Moviolia can be summed up in a few words: to use the potential of AI tools to simplify and automate a large part of the editing process. "The most time-consuming part of the editing process is the exchange with the customer, whose request is not always very precise and well formulated,” explains Nicolas Farrugia, an associate professor in the institution's Mathematical and Electrical Engineering (MEE) department and AI expert. "This frequently requires three or four round-trips, with new adjustments at each stage.” The team therefore entrusts an AI with analyzing the “brief” provided by the customer, based on the history of projects (several hundred) already carried out by Tydeo. This enables us to better understand expectations, detect potential difficulties and reformulate or “interpret” the brief.

A revolution in the corporate video market

But AI can intervene at several other levels. For example, it can perform an initial “derushing” - in other words, an analysis of all the shots - and index them. Another possibility is to work on the soundtrack, for example, transcribing speech into text - particularly for interviews. "We would like to go even further, and move on to image recognition, to automate the search for visual and audio content,” adds Nicolas Farrugia. "To use a shot with mountain images, a simple instruction would suffice to find the corresponding content in the rushes.” These are all elements that can be used to create a first edit, almost automatically. All that remains is to switch back to manual mode to refine the result. “The editor can then focus on the message and devote more time to creation,” observes Marina d'Eté, co-founder of Tydeo. All in all, AI saves a considerable amount of time - about 80% of the whole process - and therefore costs. It is enough to revolutionize the corporate video market, bringing it within reach of even the smallest structures.

The choice of a language model developed by Mistral AI

For the most part, the project uses Large Language Models (LLMs) such as those developed by Mistral AI , the French AI unicorn. "These are open models that can be easily downloaded and deployed, with no data leakage,” explains Nicolas Farrugia. "Mistral's models also have the advantage of abundant documentation, and benefit from a large community of developers. They are among the best open-access language models." As for Tydeo's database, it remains hosted in France.

The Moviolia team plans to present a first demonstrator at the end of 2025, followed by a second, more advanced one, at the end of 2026. Tydeo, for their part, have already produced a video on the project itself - a sort of mise en abyme. As far as Ektacom is concerned, they are managing the project, and preparing to increase their expertise in AI. As for the BRAIn team, with 6 researchers dedicated to Moviolia, it brings its expertise in AI, and in particular in sound signal, digital and neuroimaging... Several recruitments are underway at the three partners.
Launched in 2024, the project has a total budget of around 1.1 million euros. Moviolia is accredited by Images et Réseaux, the competitive cluster dedicated to digital innovation in the Pays de la Loire and Brittany regions.

A concern for energy-efficient AI

But Moviolia's approach goes beyond corporate videos. The project could also lead to the automated creation of a wide range of content, using the full richness of multimodal AI. Many fields could be involved - starting with scientific research. “We could, for example, produce automated transcriptions of audio or video interviews carried out in the field - by geographers or sociologists in particular”, notes Nicolas Farrugia. In the longer term, various sectors, such as the music industry or video games, could also be interested. Reflection is in full swing...
Another issue mobilizing the BRAIn team is the question of AI “frugality”. A growing number of players are concerned about the enormous energy and computing power requirements of the most advanced AIs. According to the International Energy Agency (IEA), a simple query on ChatGPT requires 2.9 watt-hours of electricity, ten times more than a Google search. And the overall electricity consumption of data centers is increasing at breakneck speed... A problem that has prompted a whole open source software community, notably the Franco-American site Hugging Face, to take action. These enthusiasts are committed to “quantifying” the various LLMs, i.e. modifying their weight to limit the amount of computation and memory required, and working to simplify their architecture to offer lighter, less power-hungry AI models. Moviolia is part of this trend: “Initially, we will be using models created by others,” explains Nicolas Farrugia. "But later on, we may try to develop our own, smaller model, specialized in video editing. In short, a more frugal model.” A way of democratizing AI. This approach can be compared with the recent emergence of Chinese AI DeepSeek, which uses far less sophisticated “chips” than Open AI for ChatGPT - with fairly convincing results.

The Moviolia team also intends to play the local economic development card. "In Finistère, few players are involved in AI research,” notes Nicolas Farrugia. "This project is an opportunity to showcase local know-how. We are trying to contribute to this movement, by developing AI and the jobs it can create in the Brest region, at a time when many young engineers are tempted to leave for Paris.”

(1) Better Representations for Artificial Intelligence.