UNIST Proposes Design Principle for More Reliable "AI Artist"

Published 22 Oct.2025 08:55(KST)

Professors Yoon Sunghwan and Yoo Jaejun's Team Paper Accepted at ICCV 2025
Revealing Training Principles to Enhance Robustness and Generalization of Diffusion Models

When you ask ChatGPT to "draw a Ghibli-style picture," it is not ChatGPT itself that creates the image, but rather an image-generating artificial intelligence called DALL·E.

Such diffusion models are capable of producing high-quality images, but sometimes generate pictures with three fingers or distorted faces. Due to their billions of parameters, it is also difficult to run them directly on devices with limited computational resources, such as smartphones.

A team of Korean researchers has proposed a design principle for generative AI that can overcome these limitations.

On October 22, the research team led by Professors Yoon Sunghwan and Yoo Jaejun at the Graduate School of Artificial Intelligence at UNIST announced that they have theoretically and experimentally demonstrated that designing diffusion models to reach flat minima of the loss function during training can simultaneously improve both the robustness and generalization performance of the models.

Research team, (from left) Professor Yoon Sunghwan, Professor Yoo Jaejun, Researcher Lee Taehwan (first author), Researcher Seo Kyungguk (first author). Provided by UNIST

Diffusion models are already being used in image-generating AI such as ChatGPT's DALL·E and Stable Diffusion because they can produce high-quality images. They can transform your photo into a Ghibli-style or Van Gogh-style painting, create a personalized character as a four-panel comic, or generate realistic images of imaginary scenes.

However, these diffusion models have a weakness in robustness, as they are vulnerable to error accumulation during the short generation process, quantization errors that occur during compression for deployment on small devices, and adversarial attacks that introduce subtle perturbations to disrupt the output.

The research team diagnosed that the robustness issue stems from the fundamental lack of generalization performance in AI. Generalization performance refers to the model's ability to operate reliably even on new data or in new environments that were not used during training.

$The mathematical proof of the relationship between flat minima of the loss function and robustness proceeds as follows. Consider a loss function $ L(\theta) $ defined over parameters $ \theta \in \mathbb{R}^n $. A minimum $ \theta^* $ is called flat if the Hessian matrix $ H = \nabla^2 L(\theta^*) $ has small eigenvalues, indicating a wide, flat basin around $ \theta^* $. **Step 1: Definition of flat minimum** At $ \theta^* $, the loss satisfies: \[ \nabla L(\theta^*) = 0, \quad \text{and} \quad \lambda_{\max}(H) \approx 0, \] where $ \lambda_{\max}(H) $ denotes the largest eigenvalue of the Hessian matrix. **Step 2: Perturbation analysis** Consider a small perturbation $ \delta $ around $ \theta^* $: \[ \theta = \theta^* + \delta. \] Using a second-order Taylor expansion of the loss around $ \theta^* $: \[ L(\theta^* + \delta) \approx L(\theta^*) + \frac{1}{2} \delta^T H \delta. \] Since $ \nabla L(\theta^*) = 0 $, the linear term vanishes. **Step 3: Flatness implies robustness** If $ \theta^* $ is a flat minimum, then the eigenvalues of $ H $ are small, so for any perturbation $ \delta $ with bounded norm \( \$

The mathematical proof of the relationship between flat minima of the loss function and robustness proceeds as follows.

Consider a loss function $ L(\theta) $ defined over parameters $ \theta \in \mathbb{R}^n $. A minimum $ \theta^* $ is called flat if the Hessian matrix $ H = \nabla^2 L(\theta^*) $ has small eigenvalues, indicating a wide, flat basin around $ \theta^* $.

**Step 1: Definition of flat minimum**

At $ \theta^* $, the loss satisfies:
\[
\nabla L(\theta^*) = 0, \quad \text{and} \quad \lambda_{\max}(H) \approx 0,
\]
where $ \lambda_{\max}(H) $ denotes the largest eigenvalue of the Hessian matrix.

**Step 2: Perturbation analysis**

Consider a small perturbation $ \delta $ around $ \theta^* $:
\[
\theta = \theta^* + \delta.
\]

Using a second-order Taylor expansion of the loss around $ \theta^* $:
\[
L(\theta^* + \delta) \approx L(\theta^*) + \frac{1}{2} \delta^T H \delta.
\]

Since $ \nabla L(\theta^*) = 0 $, the linear term vanishes.

**Step 3: Flatness implies robustness**

If $ \theta^* $ is a flat minimum, then the eigenvalues of $ H $ are small, so for any perturbation $ \delta $ with bounded norm \( \

The team found a solution to the problem in the "valley shape of the lowest point" of the loss function. The loss function numerically represents the difference between the AI's predicted results and the correct answers; the lower the value, the better the training. During training, AI seeks to minimize the loss value, but if this minimum is narrow and steep, performance can easily collapse with small perturbations. In contrast, if the minimum is wide and flat, performance remains stable even in new situations or under interference.

Among algorithms that seek flat minima, Sharpness-Aware Minimization (SAM) proved to be the most effective. Diffusion models applying SAM were less affected by error accumulation during the short generation process and experienced less quality degradation during compression (quantization) for deployment on small devices. Notably, the robustness of diffusion models was greatly improved, maintaining performance even under adversarial attacks that were seven times stronger than those used on previous models.

Originally, issues such as error accumulation from repeated generation, quantization errors during model compression, and vulnerability to adversarial attacks in diffusion models were studied separately. However, the research team explained that their study demonstrated all these issues can be addressed by the single principle of "flat minima."

The team stated, "Our work is significant in that it proposes a design principle for generative AI that can be trusted and used in various industries and real-world environments, going beyond simply improving image quality. It will serve as a foundation for enabling large-scale generative models like ChatGPT to be trained stably even with small amounts of data."

This research was co-led by Lee Taehwan and Seo Kyungguk of UNIST as first authors.

The results have been accepted for presentation at the 2025 International Conference on Computer Vision (ICCV). ICCV is one of the most prestigious conferences in the field of artificial intelligence. This year's conference opened in Hawaii, United States, on October 19 and will run for five days until October 23.

The research was supported by the Ministry of Science and ICT, the National Research Foundation of Korea, the Institute of Information & Communications Technology Planning & Evaluation, the Ministry of Health and Welfare, and Ulsan National Institute of Science and Technology.

UNIST Proposes Design Principle for More Reliable "AI Artist"

Professors Yoon Sunghwan and Yoo Jaejun's Team Paper Accepted at ICCV 2025Revealing Training Principles to Enhance Robustness and Generalization of Diffusion Models

Professors Yoon Sunghwan and Yoo Jaejun's Team Paper Accepted at ICCV 2025
Revealing Training Principles to Enhance Robustness and Generalization of Diffusion Models