Selected work

Representative Papers

Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks

Jul 08, 2026

This work investigates whether multimodal agents genuinely comprehend the physical principles and inverse problems underlying computational imaging, rather than relying solely on semantic visual capabilities. To this end, we introduce ImagingBench, the first benchmark comprising 20 tasks across five major categories, and systematically evaluate state-of-the-art vision-language models—including Gemini, GPT, and Qwen—alongside specialized non-agent methods under Expert, Planner, and Forward settings. Our results reveal that current agents consistently underperform dedicated approaches across most tasks, particularly in computation-intensive domains such as lensless imaging and event camera reconstruction. Moreover, planner-guided strategies yield only marginal and unstable improvements, exposing fundamental limitations in the agents’ ability to ensure physical consistency and achieve expert-level reconstruction fidelity.

0 citationsRead paper

The Need for Neural ISP in the Small-Pixel Era: How Shrinking Pixels Push Optics to the Limit and Neural Restoration Pushes Back

Jun 04, 2026

This work addresses the severe resolution degradation in smartphone telephoto lenses when pixel sizes shrink below 0.5 µm, where optical diffraction and geometric aberrations become dominant, and conventional image signal processors (ISPs) fail to recover fine details due to their lack of explicit point spread function (PSF) modeling. The authors propose an end-to-end trained neural ISP that explicitly compensates for residual aberrations through both single-frame and multi-frame architectures, leveraging an optical simulation platform that jointly controls signal-to-noise ratio and diffraction spot size. Experiments demonstrate that at a 0.35 µm pixel pitch, the method achieves an MTF50 of 745 cycles/mm—2.5–3× higher resolution than traditional ISPs—and reduces LPIPS to 0.151. In low-SNR multi-frame scenarios, performance approaches that of a bright single-frame baseline. This study is the first to systematically validate that neural ISPs can transform extremely small pixels into an imaging advantage, revealing that the fundamental bottleneck of conventional ISPs at small pixel scales stems from uncorrected PSF-induced blur.

0 citationsRead paper

Recent publications

Latest Papers

Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks

Jul 08, 2026

This work investigates whether multimodal agents genuinely comprehend the physical principles and inverse problems underlying computational imaging, rather than relying solely on semantic visual capabilities. To this end, we introduce ImagingBench, the first benchmark comprising 20 tasks across five major categories, and systematically evaluate state-of-the-art vision-language models—including Gemini, GPT, and Qwen—alongside specialized non-agent methods under Expert, Planner, and Forward settings. Our results reveal that current agents consistently underperform dedicated approaches across most tasks, particularly in computation-intensive domains such as lensless imaging and event camera reconstruction. Moreover, planner-guided strategies yield only marginal and unstable improvements, exposing fundamental limitations in the agents’ ability to ensure physical consistency and achieve expert-level reconstruction fidelity.

0 citationsRead paper

The Need for Neural ISP in the Small-Pixel Era: How Shrinking Pixels Push Optics to the Limit and Neural Restoration Pushes Back

Jun 04, 2026

This work addresses the severe resolution degradation in smartphone telephoto lenses when pixel sizes shrink below 0.5 µm, where optical diffraction and geometric aberrations become dominant, and conventional image signal processors (ISPs) fail to recover fine details due to their lack of explicit point spread function (PSF) modeling. The authors propose an end-to-end trained neural ISP that explicitly compensates for residual aberrations through both single-frame and multi-frame architectures, leveraging an optical simulation platform that jointly controls signal-to-noise ratio and diffraction spot size. Experiments demonstrate that at a 0.35 µm pixel pitch, the method achieves an MTF50 of 745 cycles/mm—2.5–3× higher resolution than traditional ISPs—and reduces LPIPS to 0.151. In low-SNR multi-frame scenarios, performance approaches that of a bright single-frame baseline. This study is the first to systematically validate that neural ISPs can transform extremely small pixels into an imaging advantage, revealing that the fundamental bottleneck of conventional ISPs at small pixel scales stems from uncorrected PSF-induced blur.

0 citationsRead paper