Towards Trustworthy Visual Foundation Models in the Real World

The recent surge of interest in generative AI has brought unprecedented advances in content generation, reasoning, and multimodal interaction, while also exposing a growing set of concerns that extend well beyond model performance. As generative AI systems are increasingly deployed in real-world scenarios, issues such as fairness, privacy, safety, robustness, transparency, and accountability have become central to their responsible development and use. Although substantial efforts have been made to address these challenges, existing research remains scattered across different communities and often focuses on isolated aspects of the problem. There is still a clear need for a dedicated forum that brings together methodological advances, empirical evaluations, benchmarking efforts, and deployment-oriented studies in a unified manner. This Research Topic therefore aims to consolidate emerging research on the trustworthy and responsible deployment of generative AI, with particular attention to the practical challenges and societal implications arising in real-world applications. By highlighting both technical innovations and broader deployment considerations, this article collection will help shape a more systematic and impactful research agenda for the next stage of generative AI development.Visual foundation models have rapidly advanced computer vision by enabling strong generalization across diverse perception and multimodal tasks. Models such as vision-language models and large-scale multimodal systems are increasingly deployed in real-world settings, including healthcare, autonomous systems, content moderation, robotics, and human-computer interaction. However, their deployment also raises critical trustworthiness challenges. These models may exhibit unreliable behavior under distribution shifts, hidden biases across demographic or contextual groups, privacy leakage from large-scale training data, vulnerability to adversarial manipulation, and limited transparency in decision-making. As visual foundation models move from benchmark-centric evaluation to open-world applications, it becomes essential to develop systematic methods for evaluating, aligning, and governing their behavior. This Research Topic focuses on trustworthy visual foundation models, emphasizing reliability, safety, fairness, privacy, robustness, and responsible deployment in real-world environments. This Research Topic welcomes original research, reviews, perspectives, and benchmark studies on trustworthy visual foundation models for real-world applications. Topics of interest include, but are not limited to:- reliability evaluation under distribution shifts;- robustness to adversarial or environmental perturbations;- fairness and bias analysis;- privacy-preserving learning;- safety alignment;- hallucination and uncertainty assessment;- interpretability;- responsible deployment of vision-language and multimodal foundation models.We particularly encourage submissions that bridge methodological innovation with practical deployment challenges, including new evaluation protocols, datasets, auditing frameworks, mitigation strategies, and case studies in high-impact domains. Contributions should clearly articulate how their methods or analyses improve the trustworthiness, accountability, or practical reliability of visual foundation models beyond standard benchmark performance. Topic editors: Joey Tianyi Zhou, Ismail Ben Ayed, Xin Zhang Source: https://www.frontiersin.org/research-topics/82034/towards-trustworthy-visual-foundation-models-in-the-real-world