KAIST, MIT, and Microsoft Collaborate
Development of Visual Information Restoration Technology to Enhance AI Performance
Expected to Accelerate Commercialization of Humanoid Robots and More

A visual information restoration technology that allows artificial intelligence (AI) to perceive its surroundings with greater clarity has been developed. This technology can improve GPU (graphics processing unit) memory efficiency by up to 16 times, and is expected to accelerate the commercialization of humanoid robots and on-device AI.


Comparison schematic showing the performance difference between the existing visual information restoration technology and the application of Upsample Anything. KAIST (AI-generated image)

Comparison schematic showing the performance difference between the existing visual information restoration technology and the application of Upsample Anything. KAIST (AI-generated image)

View original image

KAIST announced on the 17th that the research team led by Professor Changik Kim of the School of Electrical Engineering, in collaboration with researchers from the Massachusetts Institute of Technology (MIT) and Microsoft, has developed a universal technology called "Upsample Anything." This technology aims to enhance AI visual performance even in situations with limited memory usage.


Humanoid robots, autonomous driving systems, and AI based on world models (AI models that learn and predict the physical environment and its changes in the real world) increase computational speed and reduce memory usage by compressing input images into low-resolution feature information.


However, during the process of compressing feature information, important visual data such as small objects, thin structures, and minute defects are often lost. Conversely, processing all images in high resolution to address this issue would require enormous GPU memory and computational resources, making real-time information processing difficult.


This problem is particularly pronounced in scenarios where small devices like smartphones and highly mobile robots need to precisely recognize their surroundings. Enhancing AI's visual performance under limited memory usage is a challenge that must be overcome before the commercialization of humanoid robots and on-device AI.


In contrast, the joint research team developed Upsample Anything by restoring high-resolution feature information from low-resolution inputs using the boundary and structural information of the input images.


Conventional technologies require additional retraining or complex optimization processes before applying them to new environments and data. However, Upsample Anything can find the optimal restoration method using just a single input image, without the need for additional data training, allowing for immediate application across diverse environments.


Most notably, by compressing and utilizing only the core information, rather than storing and processing all visual data in high resolution, GPU memory usage has been dramatically reduced.


(From left) Minseok Seo, Ph.D. candidate at KAIST; Mark Hamilton, researcher at MIT and Microsoft; Changik Kim, professor at KAIST. KAIST

(From left) Minseok Seo, Ph.D. candidate at KAIST; Mark Hamilton, researcher at MIT and Microsoft; Changik Kim, professor at KAIST. KAIST

View original image

In practice, the joint research team applied Upsample Anything to images of size '224×224' (50,000 pixels), commonly used in AI research, and succeeded in restoring visual information close to the original within 0.4 seconds, improving GPU memory efficiency by up to 16 times.


This outcome demonstrates that AI can recognize its surroundings more precisely even with limited computational resources. As a result, the potential for Upsample Anything to be applied and utilized in next-generation AI fields, including smartphones, humanoid robots, autonomous driving systems, and on-device AI, is increasing.


The research was accepted as a paper at the global conference "CVPR 2026" in the fields of AI and computer vision. In addition, the team was awarded the top honor, the "CVPR Compute Gold Star," for efficient use of computational resources, and was also selected as the "Transparency Champion" for transparency and reproducibility in the research process.


Professor Kim stated, "Upsample Anything is an algorithm that can dramatically enhance AI's visual precision even with limited resources. We expect it will help accelerate the practical implementation of humanoid robots and on-device AI."



Meanwhile, Minseok Seo, a PhD candidate at KAIST, participated as the first author of this research.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing