MOSAIC - Multimodal Open models for Safe AI in Content generation and retrieval

The MOSAIC project (Multimodal Open models for Safe AI in Content generation and retrieval) aims to develop the first large-scale European multimodal foundation models for cross-modal retrieval and image/video generation, with trustworthiness and safety embedded by design. Current state-of-the-art systems such as CLIP, DINOv2, Stable Diffusion, or multimodal LLMs achieve remarkable performance but rely on uncurated web-scale data and lack robust safety mechanisms, making them prone to biased, unsafe, or misleading outputs. MOSAIC advances beyond this paradigm by combining contrastive retrieval backbones, diffusion-based generators, and contrastive deepfake detectors, while directly integrating safety mechanisms into training and adaptation procedures. 

The project will construct large-scale multimodal datasets, develop retrieval and generative models for images and videos, and fine-tune them on safe/unsafe multimodal pairs using parameter-efficient techniques such as LoRA and QLoRA. Safety is addressed through embedding adaptation, generative backbone alignment, and the development of detectors capable of reliably distinguishing synthetic from authentic content. The models will be evaluated in practical domains such as multimedia tagging, social network analysis, and news verification. 

Leveraging EuroHPC resources is essential for scaling, as training involves billions of parameters and datasets with millions of pairs, requiring distributed GPU clusters and optimized data pipelines. MOSAIC will deliver publicly available models, datasets, and tools, reinforcing Europe’s technological sovereignty, scientific leadership, and societal resilience against disinformation, while aligning with ongoing European initiatives such as ELLIOT, ELIAS, and MINERVA.