TLDR: A research team at UNIST, led by Professors Jaejun Yoo and Sung Whan Yoon, has unveiled new design principles for generative AI, specifically diffusion models, to enhance their trustworthiness, robustness, and generalization ability. By training these models to reach ‘flat minima’ in the loss landscape, they achieved reduced error accumulation, improved performance after compression, and a sevenfold increase in resistance to adversarial attacks. This unified approach offers a fundamental framework for developing versatile AI systems applicable across various industries and could lead to more efficient training of large-scale models.
ULSAN, South Korea – A groundbreaking study from the UNIST (Ulsan National Institute of Science and Technology) Graduate School of Artificial Intelligence has introduced a new paradigm for developing trustworthy generative AI systems. Professors Jaejun Yoo and Sung Whan Yoon, along with first authors Taehwan Lee and Kyeongkook Seo, have demonstrated that training diffusion models to achieve ‘flat minima’ in their loss landscape significantly improves both their robustness and generalization capabilities.
Diffusion models, which power popular AI art tools like DALL·E and Stable Diffusion, are renowned for their ability to transform images into artistic styles, create personalized characters, and render realistic landscapes. However, these models often grapple with limitations such as generating artifacts (e.g., three-fingered hands or distorted faces), accumulating errors during rapid generation tasks, degrading performance after model compression (quantization), and being vulnerable to adversarial attacks.
The UNIST team’s research, presented at the 2025 International Conference on Computer Vision (ICCV) in Hawaii (October 19-23, 2025), posits that these issues stem from a fundamental inability to generalize effectively across new or unseen data. Their solution focuses on guiding the training process towards ‘flat minima’ – broad, gentle regions in the model’s loss landscape. Unlike ‘sharp minima’ (narrow, steep valleys) which lead to performance drops with minor disturbances, flat minima ensure stable and reliable performance even with small input variations or noise.
Through extensive theoretical analysis and experiments conducted on datasets like CIFAR-10, LSUN Tower, and FFHQ, the researchers confirmed that flat minima indeed enhance both generative performance and robustness. They identified Sharpness-Aware Minimization (SAM) as the most effective algorithm for achieving this. Models trained with SAM exhibited a notable reduction in error accumulation during quick generation, maintained higher quality outputs post-compression, and demonstrated an impressive sevenfold increase in resistance to adversarial attacks.
Also Read:
- Governments Emphasize Urgent Need for Robust AI Governance Frameworks Amidst Rapid Advancements
- Generative AI Market Poised for Explosive Growth, Projected to Reach $440 Billion by 2032 Amidst Content Creation and Automation Boom
This study is particularly significant because it offers a unified solution to challenges previously addressed individually. The researchers emphasize that their findings extend beyond mere improvements in image quality, providing a foundational framework for designing versatile and trustworthy generative AI systems. This could have profound implications for various industries and real-world applications, potentially enabling more efficient training of large-scale models like ChatGPT, even with limited data. The research received support from the Korean Ministry of Science and ICT (MSIT), the National Research Foundation (NRF), the Institute for Information & Communications Technology Planning and Evaluation, and UNIST.
