Lens Blur Fields Separates Same-Model iPhones—but Needs Calibration
Lens Blur Fields remains a research method for measuring the optical behavior of individual cameras, not a service that can identify a phone from any photograph. The important update is that the work has progressed from its 2023 preprint to a 2025 journal publication: the official Lens Blur Fields project page now labels it as TPAMI 2025 and documents tests in which separate iPhone 12 Pro units produced measurably different point-spread functions.
That distinction matters because the experiment does not support the stronger claim that two ordinary photos are enough to reveal whether they came from the same handset. The method first calibrates each camera with controlled focal stacks, known patterns and a separately trained neural representation. As of this review, the paper, implementation, pretrained models and selected calibration data are public, but the repository also warns that its development code has not been maintained since 2023.
What the researchers actually demonstrated
A camera does not reproduce an ideal point of light as a perfect point. Defocus, diffraction, lens aberrations, sensor microlenses and color filters spread that light into a point-spread function, or PSF. Its shape can change across the image plane and as the camera changes focus, so a single blur kernel cannot describe the entire optical system.
Lens Blur Fields represents those changes with a multilayer perceptron. Its inputs describe image position, focus and the coordinates inside the blur kernel; an optional target-distance dimension expands a five-dimensional field into a six-dimensional one. The resulting function is specific to the calibrated camera-and-lens system rather than merely to a phone model named in metadata.
The revised July 2025 research paper reports repeatability experiments on two iPhone 12 Pro devices and two iPhone 14 Pro devices. For each phone, the team captured four focal-stack datasets for each of two repeatability conditions and trained separate models, producing eight datasets per handset. Variation among repeated estimates for one device was smaller than the root-mean-square difference between the two devices’ mean PSFs. This supports a narrower conclusion than automatic source attribution: under the researchers’ calibration procedure, the method resolved optical differences between individual phones of the same model.
Calibration is the hidden requirement
The process begins with a monitor displaying known dot grids and synthetic noise patterns while the camera is fixed in front of it. The researchers capture a stack of images over multiple focus positions, correct for perspective, lens distortion and radiometric variation, and optimize the neural network so that its predicted blur recreates the recorded images. Smartphone tests used raw captures and software-controlled focus rather than compressed photographs collected from social media.
This is why background blur in a portrait should not be treated as a ready-made serial number. A real photograph mixes the lens PSF with scene depth, motion, subject texture, computational photography, image processing and compression. In the published workflow, the known calibration pattern and controlled geometry make the non-blind deconvolution problem tractable. The authors explicitly describe blind recovery in unstructured environments as future work.
The approach is lightweight in capture equipment—a monitor and a stable camera mount are sufficient—but it is not lightweight in every computational sense. In the paper’s representative dense iPhone 12 Pro configuration, acquisition can take less than five minutes, while training the five-dimensional model took 14 hours on one NVIDIA A6000 GPU. The trained MLP occupied 19.1 MB, far below the 9.3 GB estimated for the compared discrete representation. Those figures describe that specified experimental configuration, not a guaranteed runtime for every phone.
The public release is more complete, but still experimental
The project’s current public materials add practical value that was missing from early descriptions. The official implementation repository includes preprocessing and rendering notebooks, training configurations, pretrained models for two iPhone 12 Pro wide cameras and links to released data. The listed downloads include 1.06 GB of downsampled processed data, 18.3 GB at full resolution and 1.75 GB of unprocessed calibration captures.
That release makes parts of the experiment inspectable and reproducible, but it is not a packaged consumer application. Setup requires Python, PyTorch, CUDA-oriented tiny-cuda-nn tooling and device-specific data preparation. The repository’s own maintenance warning also means users should expect dependency work rather than a polished installation path.
Why this is not yet a universal phone fingerprint
A forensic identification system needs more than evidence that two calibrated devices differ. It must estimate error rates across many units, ages, repairs and environmental conditions; distinguish genuine device variation from capture-session variation; and show how editing, resizing and compression affect matching. The paper asks how distinct the signatures remain across a wider population, which indicates that population-scale uniqueness was not established by the reported four-phone comparison.
It is also important to separate characterization from classification. Lens Blur Fields learns a detailed model for a camera whose controlled calibration images are available. The published tests compare recovered PSFs and repeatability; they do not report an operational benchmark in which unknown casual photographs are assigned to one handset among thousands.
The demonstrated value is nevertheless substantial. A compact, continuous model can reproduce spatially varying blur more faithfully than simplified analytical models, support device-specific synthetic depth of field and improve calibrated restoration experiments. Its possible forensic use is a research direction grounded in observed device differences—not yet proof that an investigator can upload one photograph and identify the physical phone that captured it.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.