Neural ISP breaks telephoto physics wall: 3x resolution at 0.35 micron

A new paper from Jingxi Li, Neerja Aggarwal, and colleagues (arXiv:2606.07675) tackles the fundamental limits of smartphone telephoto cameras as pixel pitches shrink toward sub-0.5 microns. The researchers identify a "telephoto physics wall": geometric aberrations in the lens dominate, and traditional stage-wise ISPs cannot model or invert the complex point spread function (PSF). They simulate five pixel configurations from 0.35 to 0.75 microns, scaling aperture to keep SNR and diffraction constant, isolating the effects of aberration and spatial sampling.

Their learning-based neural ISP—trained on the underlying PSF degradations—delivers dramatically better results. At 0.35 micron pitch, it achieves 745 cycles/mm MTF50, a 2.5–3x resolution jump over traditional ISPs. LPIPS (perceptual similarity) improves from 0.244 to 0.151, while conventional pipelines remain flat. In low-light multi-frame tests (15 dB SNR), the neural ISP recovers near-baseline quality, whereas traditional multi-frame ISPs show no improvement—proving that the bottleneck is PSF blur, not noise. The work suggests future telephoto modules can use simpler optics and rely on neural restoration.