Learning Lens Blur Fields

Esther Y. H. Lin1, Zhecheng Wang1, Rebecca Lin2, Daniel Miau3, Florian Kainz3, Jiawen Chen3, Xuaner (Cecilia) Zhang3, David B. Lindell1, Kiriakos N. Kutulakos1

1University of Toronto   ·   2Massachusetts Institute of Technology   ·   3Adobe Inc.

TPAMI 2025

teaser
Lens blur field shown in step 2 is from the dual pixels of a Pixel 4a.

Abstract

Optical blur is an inherent property of any lens system and is challenging to model in modern cameras because of their complex optical elements. To tackle this challenge, we introduce a high‑dimensional neural representation of blur—the lens blur field—and a practical method for acquisition.

The lens blur field is a multilayer perceptron (MLP) designed to (1) accurately capture variations of the lens 2‑D point spread function over image‑plane location, focus setting, and optionally depth; and (2) represent these variations parametrically as a single, sensor‑specific function. The representation models the combined effects of defocus, diffraction, aberration, and accounts for sensor features such as pixel color filters and pixel‑specific micro‑lenses.

We provide a first‑of‑its‑kind dataset of 5‑D blur fields—for smartphone cameras, camera bodies equipped with a variety of lenses, etc. Finally, we show that acquired 5‑D blur fields are expressive and accurate enough to reveal, for the first time, differences in optical behavior of smartphone devices of the same make and model.

Overview

Optical blur, or point spread function (PSF) is an umbrella term for a laundry list of image degrading effects such as defocus, diffraction, and aberrations. It's hard to calibrate because it varies with sensor position, focus, target distance, and where you look on the image plane. We introduce Lens Blur Fields: tiny MLPs that can model this high-dimensional PSF.

experimental setup

Our capture setup only needs a monitor + a simple phone/camera stand. The pipeline is light:

  1. Capture focal stacks of monitor patterns (in just minutes!)
  2. Solve a non-blind deconvolution to train the MLP
  3. Get a continuous, device-specific PSF model
pipeline

Applications

Two smartphones of the same make can have subtly different PSFs. We show this with the lens blur fields of two iPhone 12 Pros:

forensics

Lens blur fields let you render device-specific depth-of-field, blur a resolution chart or a 3D scene.

rendering

And with more realistic renders, we can also do better device-specific image restoration.

deblurring

Dataset

We'll be releasing the first dataset of 5D and 6D lens blur fields for smartphone & SLR lenses, plus captures used for training. Stay tuned!

dataset

BibTeX

@article{lin2025learning,
        title={Learning Lens Blur Fields},
        ISSN={1939-3539},
        url={http://dx.doi.org/10.1109/TPAMI.2025.3578587},
        DOI={10.1109/tpami.2025.3578587},
        journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
        publisher={Institute of Electrical and Electronics Engineers (IEEE)},
        author={Lin, Esther Y. H. and Wang, Zhecheng and Lin, Rebecca and Miau, Daniel and Kainz, Florian and Chen, Jiawen and Zhang, Xuaner and Lindell, David B. and Kutulakos, Kiriakos N.},
        year={2025},
        pages={1–12}}