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.

Their study was presented at the International Conference on Computer Vision (ICCV 2025), and the findings are posted on the arXiv preprint server.

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.

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.

Carbon-supported single-atom catalysts with metal-N moieties are highly promising for lithium–sulfur batteries. They can enhance redox kinetics and suppress the dissolution of lithium polysulfides. However, carbon substrate structure optimization and catalyst coordination environment modulation must be done simultaneously to maximize the potential of these catalysts.

Taking on this challenge, a team of researchers led by two associate professors from Chung-Ang University—Seung-Keun Park from the Department of Advanced Materials Engineering and Inho Nam from the Department of Chemical Engineering—has demonstrated dual‑level engineering of metal–organic framework (MOF)‑derived hierarchical porous carbon nanofibers with low‑coordinated cobalt single‑atom catalysts for high‑performance lithium–sulfur batteries. Their novel findings were published in Advanced Fiber Materials on 24 September 2025.

Dr. Park says, “Our motivation lies in addressing the fundamental materials challenges that have limited the development of next-generation energy storage systems. Lithium-ion batteries have been widely adopted but are approaching their intrinsic energy density limits.

“Lithium sulfur batteries offer much higher theoretical capacity and energy density, yet they are severely restricted by the polysulfide shuttle effect, slow redox kinetics, and rapid capacity fading. Our group has long been committed to overcoming these bottlenecks by combining structural engineering of carbon frameworks with atomic-level catalyst design.”

In this study, the researchers focused on embedding single cobalt atoms in a low-coordinated N₃ environment within a porous carbon nanofiber network. This approach enhances the adsorption of lithium polysulfides and accelerates their redox reactions, thereby mitigating the shuttle effect and improving overall kinetics. Therefore, the present work supports the belief that rational materials design at both the macro and atomic levels can solve long-standing challenges.

From a materials perspective, the proposed dual-level engineering strategy integrates a hierarchical porous carbon nanofiber structure with atomically dispersed cobalt single-atom sites in a low-coordinated N₃ configuration. The carbon nanofiber provides mechanical stability, abundant pore channels, and excellent electrolyte wettability, while the cobalt sites catalyze the adsorption and conversion of polysulfides. This synergistic design allows the battery to achieve high-capacity retention and superior rate performance over hundreds of cycles.

In the long term, the results of this study could contribute to the realization of high-performance lithium sulfur batteries for diverse real-life applications. These include electric vehicles with extended driving ranges, large-scale renewable energy storage systems that can balance intermittent solar and wind power, and lightweight, flexible power sources for portable and wearable electronics.

“Our material is free standing, binder free, and flexible. It can be directly applied as an interlayer in pouch cells and has been demonstrated to maintain mechanical integrity even under bending, while powering small devices,” points out Dr. Nam, highlighting the immense practical implications of their work.

For society, such advances mean safer and more efficient batteries that accelerate the transition to clean energy. This can reduce dependence on critical raw materials, lower costs, decrease carbon emissions, and ultimately make sustainable technologies more reliable and accessible in everyday life.

On Thursday, Tesla shareholders approved an unprecedented $1 trillion pay package for CEO Elon Musk. The full compensation plan will go into effect by 2035—assuming Musk and the company successfully hit ambitious financial and production targets. If that happens, Musk will also get control of some 25 percent of the business, up from the 12 percent he controls currently. More than 75 percent of Tesla shareholders approved the move in a preliminary vote.

Musk celebrated the news onstage at Tesla’s Gigafactory in Austin, Texas, appearing alongside two dancing humanoid robots, the company’s Optimus products. “Look at us, this is sick,” he said.

To meet its goals, however, Tesla will have to lead in industries well beyond electric cars—and guarantee that Optimus can do much more than dance. It will also have to beat all competitors in autonomous driving technology and robotics. “Tesla will have to be the market leader not just in the US but also Europe and other regions,” says Seth Goldstein, a senior equity analyst at Morningstar, a financial services firm.

Specifically, Tesla needs to hit an$8.5 trillion valuation over the next 10 years, deliver 20 million vehicles to customers, send out 1 million robots, operate 1 million robotaxis, and sell 10 million subscriptions for its “Full Self-Driving” software over a three-month period—in addition to other financial targets.

Still, the vote marks a win for Musk, whose previous package, a $50 billion payday laid out in 2018, has been caught up in litigation after a shareholder alleged that the CEO had too much influence over the company’s board and that Tesla was therefore failing to uphold its legal obligations to shareholders. The lawsuit, brought in Delaware’s Chancery Court, led to Tesla reincorporating in Texas. A panel of judges heard the case on appeal in October; they’ll likely make a final decision in the coming months.

Before the vote, Tesla’s board argued the sky-high pay package was necessary to retain Musk as CEO—and keep him focused on the car company. In a call with investors last month, Musk suggested that he would have a hard time pushing Tesla ahead in robotics and autonomy if he didn’t have a strong sway over the automaker. “If we build this robot army, do I have at least a strong influence over this robot army?” he asked. “I don’t feel comfortable building that robot army unless I have a strong influence.”

Following Thursday’s vote, Musk told investors gathered in Texas that production of the Cybercab, a self-driving vehicle that lacks a steering wheel or sideview mirrors, would begin in April. The company will need permission from the federal government to put the unconventionally designed car on the road.