kqwang/phase-recovery: a curated index of phase recovery research, not a solver

What problem a paper index solves, and for whom

Phase recovery is a naming problem before it is a maths problem. The README defines the term explicitly as calculating the phase of a light field from its amplitude or intensity measurements, and then notes that the label covers holography and interferometry, the transport of intensity equation, optimisation-based phase retrieval, wavefront sensing, and deep-learning approaches. Those communities publish in different venues, use different vocabulary, and rarely cite each other. A researcher who arrives from ptychography may not recognise that a TIE paper is solving the same inverse problem with a different measurement model.

This repository exists to collapse that vocabulary gap. It is released alongside a review in Light: Science & Applications (the DOI in the homepage field), and its structure mirrors a survey: groups by continent, companies, workshops with video or slides, then papers split into conventional and deep-learning branches. The intended reader is someone entering the field, or someone writing about it, who needs to know which names and which sub-areas exist before searching. It is not aimed at an engineer who wants a reconstruction function they can call tonight.

The paper taxonomy is the actual contribution

Most awesome-style lists are flat. This one is not, and the shape of the tree is the most informative thing in the repository. Conventional phase recovery is divided into holography and interferometry, transport of intensity equation, wavefront-sensing-based approaches, and optimisation-based approaches. That last branch is then split by the geometry of the measurements: alternating projection, axial multi-intensity, radial multi-intensity, angular multi-intensity, non-convex optimisation, and convex optimisation. The distinction between axial, radial and angular matters because it describes how the diverse intensity images are collected, which in turn determines which algorithms are even applicable.

The deep-learning branch is organised differently, by where the network sits in the pipeline rather than by measurement geometry. Pre-processing covers pixel super-resolution, noise reduction, hologram generation and autofocusing. In-processing is split into five strategies with their own abbreviations: dataset-driven network-only, physics-driven network-only, physics-connect-network, network-in-physics, and physics-in-network. Post-processing covers noise reduction, resolution enhancement, aberration correction and phase unwrapping, and a separate section covers deep learning for phase processing, meaning segmentation, classification and imaging modal transformation. That five-way split of in-processing is the part worth reading closely, because it encodes a real design question: how much of the physical forward model is retained inside the learned component versus around it.

What you actually get when you clone it

There is no installation. The contributing section instructs contributors to edit the raw file with Markdown syntax, avoid typos, not duplicate existing entries, keep formatting uniform with what is already there, and respect the ordering of additions, whether chronological or alphabetical. That is the entire contribution model: fork and pull request against a Markdown document.

There are no releases, so there is no tagged version, no changelog, and no dependency manifest. The primary language field is unpopulated, which is consistent with a repository whose content is prose and links. If you are evaluating it for a build system, the honest answer is that there is nothing to build. If you are evaluating it as a document, the relevant questions are coverage and currency, and those you can only answer by reading the sections you care about. The README is truncated in the material available here, so the completeness of the later sections, including the books, review papers and dissertations, cannot be verified from what is shown.

Navigation aids and the search strategy the README recommends

The README does not expect you to read it top to bottom. It provides a table of contents with anchor links, and for the people and groups section it gives an explicit search instruction: use Ctrl + F with a keyword list. That list is itself a compact definition of the field's scope: phase imaging, holography, interferometry, phase retrieval, Fourier ptychography, inverse problem, transport of intensity equation, wavefront sensing, adaptive optics, phase unwrapping, fringe analysis, coherent diffractive imaging, optical diffraction tomography, computational imaging, biomedical imaging.

This is a small but real usability decision. A list of several hundred names is useless if you cannot filter it, and telling the reader which strings to search for is more useful than hoping they guess the right synonym. The groups are organised by continent, with Asia, the Americas, Europe and Oceania as separate anchors, and each entry pairs a person or pair of names with an institution and a keyword line. The keyword line is the useful part, since it tells you whether a group works on TIE or on ptychography before you follow the link. Note that the README points to a second repository, kqwang/computational-imaging, for additional groups, so this list is explicitly not exhaustive on its own.

The limitation: a reading list cannot reconstruct anything

The failure mode is straightforward. If you arrive expecting an implementation of alternating projection, a TIE solver, or a trained network for phase unwrapping, you will find none of it. The repository contains references to work that does those things, not the work itself. There is no example dataset, no evaluation script, no metric, and no baseline. Any performance comparison you want to make has to be assembled from the cited papers, each with its own experimental conditions.

There is a second, quieter limitation. A curated index decays. Entries are added by pull request, and the README sets formatting and ordering conventions but describes no review process, no update cadence, and no criteria for inclusion beyond the contributor guidelines. The last push date is 2026-06-11, which tells you the repository has been touched recently but not which sections were touched. For a fast-moving sub-area such as deep-learning-based phase recovery, where the in-processing taxonomy already carries five competing strategy labels, a list that is accurate at the time of the review can drift within a year. Treat the deep-learning sections as a snapshot of how the field was organised when the review was written, not as a live feed.

Alternatives and how they differ in kind

The closest alternative is the companion repository, kqwang/computational-imaging, which the README itself links for additional groups. The difference is scope rather than method: computational imaging is the parent field, so that list is broader and shallower on phase-specific topics, while this one goes deep on the phase recovery sub-tree, including the measurement-geometry split and the deep-learning pipeline split that the broader list has no reason to carry. If your question is which group works on optical diffraction tomography, the broader list may be the better starting point. If your question is which papers combine a physics model with a network, this one is organised to answer it.

A different kind of alternative is a software package from one of the listed groups. That is not a like-for-like swap: a package gives you a specific algorithm with a specific interface and a specific set of assumptions about the optical setup, whereas this repository gives you the set of assumptions across many setups and lets you choose. The trade is depth for breadth. A package can be run and compared; this repository can only be read. Neither substitutes for the other, and the repository does not pretend to.

Licence, maintenance and the cost of contributing

The repository is MIT licensed. For a document of links and names, that is permissive in the usual way: you can copy, adapt and redistribute the compilation, subject to the licence terms. It says nothing about the licences of the linked papers, datasets or group websites, which are governed by their own terms, and nothing here should be read as legal advice on reuse of third-party figures or text. If you plan to reuse the taxonomy itself in a survey or course, the MIT grant is what you would rely on.

Maintenance cost is low and mostly social. There is no build to break and no dependency to upgrade. The ongoing work is keeping entries accurate and resisting duplication, which the contributing guidelines address directly with the instruction not to add what is already there and to match the existing format. The practical risk is not technical debt but editorial drift: a list maintained by drive-by pull requests can accumulate dead links and stale affiliations without anyone noticing. The repository does not appear to ship a link checker or any automated validation, based on the material available, so verifying that a given entry still resolves is the reader's job.

Editorial conclusion

Adopt this repository if you are scoping a phase recovery problem and need a structured entry point into the literature, or if you are preparing a related review and want to see how one group organised the field. Do not adopt it if you need runnable reconstruction code, reference datasets, or benchmark numbers; the repository supplies none of these, and the absence of releases means there is no versioned artefact to pin. Before relying on it, verify that the paper categories you need are actually populated rather than listed as headings, and check the companion computational-imaging repository for the groups that are missing from this one.

Official sources

1. Issues
2. kqwang/phase-recovery on GitHub
3. License: MIT
4. Project website
5. README