Vicky Kalogeiton awarded an ERC Starting Grant for FLASH, advancing more resource-efficient generative AI
We are pleased to highlight that Vicky Kalogeiton, researcher at the Computer Science Laboratory (LIX) at École Polytechnique, and a Hi! PARIS Fellow, has been awarded a prestigious European Research Council (ERC) Starting Grant for her project FLASH – From scaling to efficiency laws for visual synthesis.
The project addresses one of the major challenges facing today’s artificial intelligence systems: how can we develop high-performing generative AI models while using fewer computational and data resources?
From Scaling to Efficiency
In recent years, progress in generative AI has largely relied on increasingly large models trained on massive datasets and requiring substantial computing resources.
With FLASH, Vicky Kalogeiton aims to explore a different path: moving from a paradigm based primarily on scale toward one centered on efficiency.
Her research will focus on three key dimensions:
- Training data: understanding how much data is really necessary to achieve strong performance and how the diversity and quality of that data affect results.
- AI models: developing approaches that make better use of the information contained in training data to guide model learning more efficiently.
- Human-AI interactions: reducing the number of exchanges required between users and generative AI systems to obtain the desired result.
Preliminary research by Vicky Kalogeiton and her collaborators has already shown that it is possible to use up to 1,000 times less training data while maintaining strong performance in text-to-image generation.
Making generative AI more accessible and sustainable
The ambition of FLASH goes beyond reducing computational costs.
More efficient models could help make generative AI more accessible, sustainable and easier to deploy, including on devices with limited computing capacity.
The project will investigate ways for AI systems to interact more efficiently with users. Rather than requiring repeated prompts, for example, a model could ask clarifying questions before generating an image or first provide a preliminary sketch.
Such approaches could reduce unnecessary computations while also improving the user experience.
From visual generation to robotics
FLASH will primarily investigate image and video generation, but its methods will also be explored in the field of embodied AI and robotics.
In these environments, efficiency is particularly important. A robot must understand its surroundings, process visual information and determine an appropriate sequence of actions while operating under constraints in terms of data and computational resources.
The project could also contribute to the development of more efficient embedded AI models, capable of running directly on mobile devices rather than relying systematically on remote servers. This could bring additional benefits in terms of privacy, accessibility and energy consumption.
Supporting frontier research in AI
ERC Starting Grants support outstanding early-career researchers pursuing ambitious and innovative research projects.
Through FLASH, Vicky Kalogeiton will explore new foundations for visual generative AI, with the ambition of demonstrating that future AI systems can combine high performance with greater efficiency and sustainability.
Vicky Kalogeiton conducts her research at LIX, a joint research unit of CNRS, École Polytechnique and Institut Polytechnique de Paris.