CVPR Findings 2026  ·  Computational Imaging

Adaptive Continuous Kernel Networks for Image Reconstruction from Non-Uniform Sampling

Camille Biscarrat1,  Michaël Gharbi2,  Rahul Goel1,  Jonathan Ragan-Kelley1,  Fredo Durand1,  Tzu-Mao Li3

1MIT CSAIL    2Reve    3UC San Diego

PDF (Large) PDF (Small) Supplementary Code Image Viewer
Burst reconstruction comparison on a real scene

Our method demosaicks, denoises and merges raw burst frames into a clean output image.

Abstract

Most deep learning image enhancement and reconstruction algorithms are restricted to images represented as square lattices of pixels. This becomes insufficient in several scenarios: (1) multiple frames with subpixel alignment, such as reconstructing an image from a noisy burst; (2) radial distortions; and (3) lateral chromatic aberration where the per-wavelength warp leads to sample locations that are not only non-integer but different for the three color channels, violating the assumption of all demosaicking methods. We enable deep learning for image reconstruction with input samples at non-integer locations. We use subpixel sample location information and learn continuous reconstruction kernels, thereby maximally preserving information and avoiding degradation from resampling. The kernels are represented using neural networks conditioned on sample location as well as image information obtained from a coarser image reconstruction. Our model successfully demosaicks, denoises and merges stacks of burst frames across varying noise levels. We also demonstrate how this method can correct for chromatic aberrations in single images, making it, to our knowledge, the first joint denoising, demosaicking and chromatic aberration correction.

Method

Method overview

Our approach reconstructs a clean, demosaicked image starting from a noisy burst of misaligned raw frames, or misaligned channels from the same image in the case of chromatic aberrations (main figure) . The encoder takes the naively dmeosaicked frames and outputs per-pixel, per-channel latent codes. The MLP kernel network uses the float aligned sample locations and latent codes to output a weight for each sample in the neighborhood of the output pixel. The final pixel value is a weighted average of the neighboring samples. Both encoder and kernel networks are trained end-to-end.

BibTeX

@inproceedings{biscarrat2026ackn,
  title     = {Adaptive Continuous Kernel Networks for Image Reconstruction from Non-Uniform Sampling},
  author    = {Biscarrat, Camille and Gharbi, Micha{\"e}l and Goel, Rahul and Ragan-Kelley, Jonathan and Durand, Fredo and Li, Tzu-Mao},
  booktitle = {CVPR Findings},
  year      = {2026},
}