GAN Inversion survey: a reading list organised by what you want to do with the latent code

A survey companion, not a code repository

The repository describes itself in one sentence: a collection of resources on GAN inversion, as a supplement for the survey. The survey itself is titled GAN Inversion: A Survey, published in IEEE Transactions on Pattern Analysis and Machine Intelligence, and the arXiv entry 2101.05278 is the primary link on the page. There is also a project page, and the authors are Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou and Ming-Hsuan Yang.

The repository layout confirms what kind of artefact this is. There are four files at the root, LICENSE, README.md, reference.bib and save_cited_bib.py. The README is the list itself, the BibTeX file holds the citations, and the Python script is a helper for producing the bibliography in markdown format. There is no package, no build system, no test suite and no source tree.

That matters for how you use it. The primary language of the repository is TeX, which is a slightly odd choice for a markdown list and reflects the paper-first nature of the project rather than any software component.

The list is maintained collaboratively. The README asks for pull requests if you find work missing or have suggestions, covering papers, implementations and other resources, and it says you could manually edit items or use the script from a separate arxiv_daily_tools repository to produce them in the markdown format. So the update path is a markdown edit or a tool-assisted one, and the citation helper turns the BibTeX entries into the rendered list.

The project is MIT licensed and the last push was on 2026-07-14, with no GitHub releases. For a list, that is exactly the right shape, and it is worth saying plainly at the outset that nothing here is software you run.

The table of contents is the real contribution

The README hides its table of contents inside a details element, which is a poor decision for a document whose value is navigation, but the structure it contains is the reason to use the list at all.

The first level separates Inverted Pretrained Models from GAN Inversion Methods. The first asks which generators people have taken and fitted, and it splits into 2D GANs and 3D GANs. The second asks how a given image is mapped into a generator's latent space, and it splits into 3D GAN Inversion and 2D GAN Inversion.

Then come the operation-oriented sections, Diffusion Inversion, GAN Latent Space Editing and Diffusion Latent Space Editing. The deliberate inclusion of diffusion alongside GANs in all three is the clearest signal of how the field moved after the survey was written, since inversion and latent editing are now treated as parallel problems in two model families rather than as a GAN-specific technique.

The last major block is Applications, split into Image and Video Generation and Manipulation, Image Restoration, Image Understanding, Face Recognition, 3D Reconstruction, and an Other Applications subsection that itself contains Compressed Sensing, Medical Imaging, and Compression, Fairness, and Security.

Two things follow from that shape. First, the list is organised by the task you are trying to do, not by the year a paper appeared, so a reader looking for inversion methods for medical imaging does not have to know whether the relevant work is from 2021 or 2024. Second, the compression, fairness and security grouping shows that the survey authors treated inversion as a component with consequences outside image generation, which is the framing that makes the list useful to someone who is not building a GAN application at all.

What the 2D GAN entries reveal about what the survey considered settled

The visible part of the list, the 2D GANs section, is a good sample of how the curators judge relevance, and it is worth reading as a statement about the field rather than as a reading list.

The entries are given with authors, a venue, and links to a PDF, a project page and, where one exists, a code repository. Scaling up GANs for Text-to-Image Synthesis is a CVPR 2023 highlight, StyleGAN-T for fast large-scale text-to-image synthesis is ICML 2023, StyleGAN-XL for scaling to large diverse datasets is SIGGRAPH 2022, and Self-Distilled StyleGAN is also SIGGRAPH 2022. Ensembling Off-the-shelf Models for GAN Training is CVPR 2022.

Then the StyleGAN line itself, going backwards: StyleGAN3, Alias-Free Generative Adversarial Networks, NeurIPS 2021, with both the NVIDIA repository and a PyTorch port by Rosinality linked; StyleGAN2-Ada for training with limited data, NeurIPS 2020, with a Steam StyleGAN2-ADA entry; and StyleGAN2, whose title is cut off in the list as Analyzing and Improving the Image Quality of S...

The pattern is that the newer entries are about scale, diversity and data rather than about image quality per se. Two of them are explicitly about text-to-image synthesis, one is about scaling to large diverse datasets, one is about generating from internet photos, one is about training on limited data, and one is about ensembling separately trained models for training. That reads as the survey's position that the generative quality questions were largely answered by the time of writing, and that the open problems had moved to how to get more data, more scale and more text conditioning.

The StyleGAN entries are also the ones with the most complete linking, PDF plus project plus code, which is a useful signal for a reader deciding where to start.

Keeping a citation list honest is the unglamorous part of the work

The bibliography handling is a small detail that shows how seriously the project is taken, and it is the part a reader is most likely to copy.

The README embeds a BibTeX entry for the survey itself in a collapsible citation block, with the full author list, the journal, and the year. The repository root also holds a reference.bib, and the script save_cited_bib.py exists to generate the citation entries.

The reason the script is needed is that the markdown list links out to PDFs, project pages and code repositories for every paper, and a list with many entries accumulates citation metadata by hand otherwise. The alternative mentioned in the README is a separate arxiv_daily_tools repository, which suggests the list is partly assembled by scraping and then curating, and the script converts what comes out of that into the format the README expects.

That workflow has a consequence for the reader. The list is a curated pointer set, not a checked bibliography. An entry existing in the list means a curator judged it relevant and recorded its links, and it does not mean the venue, the year or the author list was verified against the published version. The venue years in the list are given next to each entry, which is helpful, but a citation that matters for your own paper should be taken from the paper itself rather than from the list.

There is also a licensing question worth one sentence. The repository is MIT licensed, and the list is a collection of pointers to other people's papers and code, so the MIT licence covers the list itself and not the linked works.

The diffusion sections are the reason the list is still worth opening

The most consequential structural decision in the list is the parallel treatment of diffusion models.

There is a Diffusion Inversion section next to the GAN inversion sections, a Diffusion Latent Space Editing section next to GAN Latent Space Editing, and the 3D and 2D splits in the inverted-models part are mirrored by the fact that both model families get their own treatment in the method and editing parts. The application sections are shared, which is the interesting part, since an application of latent space editing does not care which family produced the latent space.

The reason this matters for a reader arriving now is that the mental model the survey encodes, inversion as a special property of GANs, is a historical one. Once diffusion models have a workable notion of inverting a sample into a noise prediction, the two problems line up: both are asking how to move between an observation and a representation you can edit. A list that separates them by model family is organised around a distinction that carries less and less information, and a list that separates them by task is the more durable structure.

The list here does both, which is the sensible answer. The Applications block is organised by task and covers both families, and the Diffusion Inversion and Diffusion Latent Space Editing sections exist as their own entries. A reader can enter through either route.

One caveat about recency. The last push was on 2026-07-14, and the survey was published in 2022 with an arXiv entry from January 2021, so the underlying taxonomy predates the work of the last several years while the list has been kept up to a more recent date. That combination is the normal situation for a maintained companion to a fixed paper, and it means the paper is the authoritative account of the framework while the list is the more current index.

Where this repository is the wrong resource

The limits here are obvious once you know what the artefact is, and they are worth stating so nobody downloads it expecting a toolkit.

The first is that there is no implementation here. If your plan was to find out how to invert a latent code for your own model, this list will point you at papers and at other people's code repositories, and every one of those code repositories has its own dependencies, its own framework version and its own maintenance state. The list is a directory, and directories do not run.

The second is that it is a fixed taxonomy. A survey's section structure is a snapshot of how the field was organised when the survey was written, and adding a new section later would misrepresent the paper. So the list will not reflect the current state of the field the way a fresh reading of the last two years of papers would, and its fastest growing sections, the ones about 3D-aware GANs and about diffusion inversion, are the ones most likely to be incomplete relative to a 2026 perspective.

The third is the depth per entry. Each item is a title, an author list, a venue, and links. There is no annotation of what a method assumes, what it needs, or how it compares to its neighbours, which means the list saves you the step of finding the papers but not the step of deciding which to read. For a field with a few hundred inversion papers, that is the step that takes the time.

The fourth is the practical mechanics of using it. The list lives in a README on GitHub, so a search engine will find individual entries but not the structure, and the table of contents is collapsed by default inside a details element, which hides the structure from anyone who lands on the page without expanding it. For a document whose main contribution is organisation, that is an unfortunate interface choice, and the practical workaround is to read the paper's own taxonomy, which is the version with the figures and the discussion.

The fifth is that the interesting entry is a paper. If you want the survey itself, the arXiv PDF and the TPAMI version are the artefacts to read, and the repository is a companion. It is a good companion, and for the task of finding the right paper in a specific application area it is genuinely useful, but it is not a substitute for the thing it supplements.

Editorial conclusion

This repository is worth bookmarking for a researcher who is entering GAN inversion from a specific task rather than from the theory, because the organisation by application area, by model family and by editing operation is a better entry point than a chronological paper list. It is the wrong resource for anyone looking for a working implementation, since the repository contains no code beyond a bibliography helper, and the wrong resource for a citation you need to trust, since the entries are pointers to the original papers rather than to a checked bibliography. Before relying on a list entry, follow the link to the paper or the project page, check the venue year against the survey's own coverage, and remember that the last push was on 2026-07-14, so the list reflects the field as of that date rather than as of today.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. weihaox/GAN-Inversion on GitHub