Artificial intelligence is increasingly being used to see and understand the world around us.
From facial recognition on smartphones to self-driving cars and humanoid robots, computer vision has become one of the most important technologies in modern AI.
Now, researchers from the Korea Advanced Institute of Science and Technology (KAIST), working with scientists from the Massachusetts Institute of Technology (MIT) and Microsoft, have developed a new technology that allows AI systems to see more clearly while using much less computer memory.
The breakthrough could help accelerate the development of humanoid robots and powerful AI systems that run directly on devices such as smartphones.
The new method, called “Upsample Anything,” was recently accepted for presentation at the prestigious Conference on Computer Vision and Pattern Recognition (CVPR) 2026.
It also received awards for its efficient use of computing resources and its transparent research practices.
One of the biggest challenges in AI vision is balancing image quality and computing power. Modern AI systems often shrink images into lower-resolution versions before processing them. This reduces memory use and speeds up calculations.
However, this shortcut comes with a cost. Important details can disappear during compression. Small objects, fine structures and tiny defects may be lost, making it harder for AI systems to understand what they are seeing.
The alternative is to process everything at high resolution. But this requires enormous amounts of graphics processing unit (GPU) memory and computing power, making real-time operation difficult, especially for small devices and mobile robots.
This problem is particularly important for technologies such as autonomous vehicles and humanoid robots.
A robot may need to identify and pick up small objects, while a self-driving car must quickly recognize people, road signs and obstacles. These systems need both accurate vision and efficient computing.
To solve this challenge, the research team developed a new way to restore image details without requiring additional training.
Most existing methods need large amounts of new training data or complicated optimization procedures before they can work in different situations.
In contrast, Upsample Anything can determine the best way to reconstruct details using only a single input image. This means the technology can be applied immediately to many different environments.
The system works by storing only the most important visual information and then rebuilding missing details by using the edges and structures found in the original image. In simple terms, it fills in the missing information intelligently rather than processing every pixel at full resolution from the beginning.
Using a standard image size commonly employed in AI research, the system reconstructed images that were very close to the originals in only about 0.4 seconds. At the same time, it improved GPU memory efficiency by up to 16 times.
The researchers say this technology could make AI systems both smarter and more practical. By reducing memory demands while maintaining high visual accuracy, the method could enable more advanced AI features on smartphones, improve the capabilities of humanoid robots and help autonomous vehicles better understand their surroundings.
The team believes that giving AI sharper vision with fewer resources may play an important role in bringing the next generation of intelligent machines into everyday life.
