UNIST Unveils Design Principles for Trustworthy AI Art

Abstract

Flat minima, known to enhance generalization and robustness in supervised learning, remain largely unexplored in generative models. In this work, we systematically investigate the role of loss surface flatness in generative models, both theoretically and empirically, with a particular focus on diffusion models. We establish a theoretical claim that flatter minima improve robustness against perturbations in target prior distributions, leading to benefits such as reduced exposure bias — where errors in noise estimation accumulate over iterations — and significantly improved resilience to model quantization, preserving generative performance even under strong quantization constraints. We further observe that Sharpness-Aware Minimization (SAM), which explicitly controls the degree of flatness, effectively enhances flatness in diffusion models even surpassing the indirectly promoting flatness methods — Input Perturbation (IP) which enforces the Lipschitz condition, ensembling-based approach like Stochastic Weight Averaging (SWA) and Exponential Moving Average (EMA) — are less effective. Through extensive experiments on CIFAR-10, LSUN Tower, and FFHQ, we demonstrate that flat minima in diffusion models indeed improve not only generative performance but also robustness.

When users ask ChatGPT to generate an image in a Ghibli style, the actual image is created by DALL·E, a tool powered by diffusion models. Although these models produce stunning images-such as transforming photos into artistic styles, creating personalized characters, or rendering realistic landscapes-they also face certain limitations. These include occasional errors, like three-fingered hands or distorted faces, and challenges in running on devices with limited computational resources, like smartphones, due to their massive number of parameters.

A research team, jointly led by Professors Jaejun Yoo and Sung Whan Yoon of the UNIST Graduate School of Artificial Intelligence at UNIST, has proposed a new design principle for generative AI that addresses these issues. They have shown, through both theoretical analysis and extensive experiments, that training diffusion models to reach 'flat minima'-a specific type of optimal point on the loss surface-can simultaneously improve both the robustness and the generalization ability of these models.

Diffusion models are widely used in popular AI applications, including tools like DALL·E and Stable Diffusion, enabling a range of tasks from style transfer and cartoon creation to realistic scene rendering. However, deploying these models often leads to challenges, such as error accumulation during short generation cycles, performance degradation after model compression techniques like quantization, and vulnerability to adversarial attacks-small, malicious input perturbations designed to deceive the models.

The research team identified that these issues stem from fundamental limitations in the models' ability to generalize-meaning their capacity to perform reliably on new, unseen data or in unfamiliar environments.

Figure 3. A conceptual illustration of theoretical analysis. Theorem 1 (Corollary 1 for diffusion model) translates the perturbation in the parameter space into the set of perturbed distributions. Theorem 2 (Corollary 2 for diffusion model) shows that flat minima lead to robustness against the distribution gap.

To address this, the research team proposed guiding the training process toward 'flat minima'-regions in the model's loss landscape characterized by broad, gentle surfaces. Such minima help the model maintain stable and reliable performance despite small disturbances or noise. Conversely, 'sharp minima'-narrow, steep valleys-tend to cause performance to deteriorate when faced with variations or attacks.

Among various algorithms designed to find flat minima, the team identified Sharpness-Aware Minimization (SAM) as the most effective. Models trained with SAM demonstrated reduced error accumulation during rapid generation tasks, maintained higher quality outputs after compression, and exhibited a sevenfold increase in resistance to adversarial attacks, significantly boosting their robustness.

While previous research addressed issues like error accumulation, quantization errors, and adversarial vulnerabilities separately, this study shows that focusing on flat minima offers a unified and fundamental solution to all these challenges.

The researchers highlight that their findings go beyond simply improving image quality. They provide a fundamental framework for designing trustworthy, versatile generative AI systems that can be effectively applied across various industries and real-world scenarios. Additionally, this approach could pave the way for training large-scale models like ChatGPT more efficiently, even with limited data.

The research was led by first authors Taehwan Lee and Kyeongkook Seo of UNIST. Their findings have been accepted for presentation at the 2025 International Conference on Computer Vision (ICCV), one of the most prestigious forums in AI research, took place in Hawaii from October 19-23, 2025.

This study was supported by the Korean Ministry of Science and ICT (MSIT), the National Research Foundation (NRF), the Institute for Information & Communications Technology Planning & Evaluation, and UNIST.

Journal Reference

Taehwan Lee, Kyeongkook Seo, Jaejun Yoo, and Sung Whan Yoon, "Understanding Flatness in Generative Models: Its Role and Benefits," ICCV '25, (2025).