KAUST’s eye-inspired imaging approach wins Best Paper Award at SIGGRAPH Asia 2025

About

For decades, the quest to make cameras smaller and sharper has followed a familiar path: add more lenses, correct more aberrations and accept growing optical complexity as the cost of high image quality. As imaging systems continue to shrink, however, that strategy is approaching physical and practical limits.

Researchers at King Abdullah University of Science and Technology (KAUST) have now demonstrated a different approach inspired by the human eye. Instead of correcting optical imperfections across the entire image at once, their system corrects aberrations only where needed and reconstructs the rest computationally. By focusing on one small region at a time, similar to the eye’s fovea, and combining multiple views, the researchers show that it is possible to recover a full, high-quality image using far simpler optics, specifically a single achromatic doublet lens with a dynamic optical element rather than the complex compound lenses found in modern cameras.

The work, Fovea Stacking: Imaging with Dynamic Localized Aberration Correction, earned the Best Paper Award at ACM SIGGRAPH Asia 2025, one of the premier international conferences in computer graphics and interactive techniques.

“Receiving this award is especially meaningful because it recognizes a shift in how imaging systems can be designed,” said Wolfgang Heidrich, professor of computer science at KAUST and senior author of the study. “It reflects the team’s effort to rethink how image formation can be shared between optics and computation.”

Correcting vision where it matters most

Conventional cameras rely on carefully engineered stacks of lenses to suppress aberrations across the entire field of view. These designs deliver high image quality but come with trade-offs in size, weight and manufacturing complexity. When optics are simplified, image quality typically degrades, particularly toward the edges of the frame, where aberrations become difficult to correct after capture.

“At a fundamental level, we asked whether a camera really needs to be perfect everywhere at once,” said Shi Mao, a Ph.D. candidate at KAUST and lead author of the study.“Human vision does not work that way. We see sharply in a small region and build a complete picture by moving our gaze.”

At the center of the system is a deformable phase plate (DPP), a thin optofluidic optical element designed to be reshaped electronically. Instead of correcting aberrations globally, the phase plate is tuned to correct distortions only within a localized region of interest, the fovea. Unlike mechanical systems that must physically rotate a lens to change viewing direction, the DPP shifts this region of interest electronically using electrostatic actuation. This allows the system to scan the scene without the vibrations or wear associated with moving parts.

Each individual image remains partially blurred. When the sharp regions from multiple images are combined, however, the system reconstructs a composite image that is largely free from the aberrations introduced by the simplified optical design. In practice, the researchers show that as few as three to five images are sufficient to recover image quality across the full field of view.

Bridging optics and computation

Achieving this level of control requires more than hardware alone. The deformable phase plate responds nonlinearly to electrical actuation, making traditional control models inaccurate, particularly for larger corrections. To address this challenge, the researchers developed a neural network-based control model that learns how to translate desired optical wavefronts into physically achievable control signals.

“The learning-based model allows us to close the gap between what we design computationally and what the hardware can actually produce,” said Yogeshwar Nath Mishra, a co-author of the study.

The team combined this control strategy with differentiable optical modeling, allowing them to jointly optimize the optical deformation patterns correction. As a result, the system determines how the phase plate should deform to efficiently cover the full field of view with limited saccades. 

Although the approach shares similarities with traditional focus stacking, it addresses a broader class of optical distortions. In addition to extending depth of field, Fovea Stacking corrects off-axis aberrations that typically degrade image quality toward the edges of the frame. The researchers also demonstrated foveated video, dynamically tracking a moving object and keeping it within the corrected region in real time.

Rather than replacing optics with software, the researchers describe Fovea Stacking as a form of co-design, in which optical components and computational methods are developed together. As deformable optical elements continue to mature and become more compact, the approach could inform future camera designs for mobile devices, scientific instruments and foveated virtual reality displays.

More broadly, the work highlights a shift in how cameras are conceived, not as static optical pipelines, but as adaptive systems that actively decide where to look, how to correct and how to combine information over time. In doing so, it suggests that future cameras may see the world less like a traditional lens and more like the human eye.

Reference

Mao, S., Mishra, Y. N., & Heidrich, W. Fovea Stacking: Imaging with Dynamic Localized Aberration Correction. Proceedings of ACM SIGGRAPH Asia 2025. | article,. website.