# Nineteen papers chosen from 2,359 acceptances, generated by a script > top-cvpr-2023-papers is a nineteen-row curated table of CVPR 2023 papers, each with a code repository and an arXiv link, grouped under loose topic labels, auto-generated from an automation/ folder and released under CC0 with four sibling repositories for the following years. **SkalskiP/top-cvpr-2023-papers** — This repository is a curated collection of the most exciting and influential CVPR 2023 papers. 🔥 [Paper + Code] - Repository: https://github.com/SkalskiP/top-cvpr-2023-papers - Stars: 650 · Forks: 61 - Language: Python - License: CC0-1.0 - Published: 2026-09-17 · Updated: 2026-09-17 · Language: en - Canonical page: https://hysenlabs.com/projects/skalskip-top-cvpr-2023-papers ## The table is generated, and the warning points at a missing file The papers table is not typed by hand. It sits between two identical marker comments, AUTOGENERATED_COURSES_TABLE, and carries its own warning: do not edit this table manually, it is automatically generated, and head over to CONTRIBUTING.md for details on how to make changes properly. The root of the repository contains four entries: .gitignore, LICENSE, README.md and automation/. So the generator lives in automation/, which is also why the primary language is Python, and CONTRIBUTING.md is not among the top-level files. That is the one broken promise in an otherwise tidy repository. The instruction to edit through a contribution guide points at a document the tree does not contain, so anyone who wants to add a paper has to read the automation script and work out the expected entry format themselves. Which is a reasonable thing for a maintainer to fix, and a poor first experience for a contributor. For a reader it does not matter at all, since the generated output is correct and formatted. It matters for the project if the shortlist is meant to grow. ## Nineteen rows against 2,359 accepted papers The scale is stated up front and it is worth holding in mind while reading the table. CVPR 2023 received 9,155 submissions and accepted 2,359 of them. The shortlist here has nineteen rows, and the README's own advice for anything missing is to go to the full accepted-paper list on the conference site. So this is a filter, not a ranking. There is no score column, no ordering by citation or by award, and no statement of what the selection criteria were beyond the author's judgement. Nothing in the table claims to be the best paper of the year, and the entries span research groups, benchmarks and method papers alike. That framing makes the repository useful for a reader with a topic in mind, since the topic column narrows nineteen entries faster than it would narrow 2,359, and useless as a substitute for reading the conference programme. If you need the actual decision on which paper won an award, this list will not tell you, because it does not claim to track awards. ## One topic label per row, at three different granularities The taxonomy is the loose part. Labels in the visible table include Segmentation, Segmentation and Generative AI, NeRF, 3D, Generative AI, Transfer Learning and Benchmarking. Those sit at three different levels. A method class such as Segmentation is broad, a technique such as NeRF is narrow, and a combined label such as Segmentation and Generative AI is an editorial judgement about one specific paper. Nothing states a rule for choosing between them, so the same paper could plausibly have been filed two ways. The grouping is also not strict. Segmentation opens the table, then the rows move through NeRF, 3D, Generative AI, 3D, Transfer Learning, and then back to Segmentation for an open-vocabulary entry near the middle, before continuing with Generative AI and Benchmarking. So the topic column is usable as a filter but not as an index into contiguous blocks. For a script that generates this table, that is a small design choice with a real effect: readers who scan for a topic will find it, but readers who assume papers of one topic are adjacent will be wrong twice in the first dozen rows. ## Two links per row, and the second one is usually a preprint Every row carries a repository link and a paper link, side by side, and the repository comes first. That pairing is the reason to keep the table at all: a paper without code is inert for most readers, and a repository without the paper is hard to evaluate. The paper links are arXiv identifiers rather than publisher pages, and reading the identifiers is informative. Many of them are late 2022 numbers, in the 2207, 2208, 2211 and 2212 series, with others in early 2023. Those are the months when the preprints circulated, before the conference, so the list points at the version that existed while the work was still being reviewed. The practical consequence is that the linked PDF is not always the camera-ready version. Page numbering, author lists and occasionally method details differ between a preprint and the published paper, and anyone preparing a citation or comparing against a table in the proceedings should open the final version too. A handful of rows are benchmarks or measurement papers rather than methods, and those still carry a repository, which is consistent with the pattern: the code artefact is the unit of curation. ## Corporate labs and single-author repositories share one column The authorship of the code behind these papers is strikingly mixed. Large organisations appear several times: facebookresearch for PACO, open-vocabulary segmentation with mask-adapted CLIP, and Omni3D, and google for DynIBaR and DreamBooth. microsoft is there for X-Decoder, NVlabs for ODISE, SamsungLabs for SPIn-NeRF, baaivision for Painter, MadryLab for the transfer learning data work, and SHI-Labs for OneFormer. Against those sit repositories that look like one person's project: MoyGcc for Vid2Avatar, dunbar12138 for 3D-aware conditional image synthesis, ShirleyMaxx for virtual marker human mesh, timothybrooks for InstructPix2Pix, yu-takagi for brain-activity image reconstruction, rohitrango for the beyond-mAP evaluation work, YuliangXiu for ECON and VITA-Group for NeuralLift-360. The column treats them identically, which is the right editorial call for a shortlist and also the reason a reader cannot infer anything about effort or provenance from the table. Eight or nine rows are individual efforts, and a repository's popularity says nothing about whether its paper belongs in a list. ## One year per repository, four years of the same table The project is a series rather than a single repository. The header links four siblings: the 2023 list you are reading, plus 2024, 2025 and 2026 versions under the same owner, each with the same generated table structure. That design choice keeps each year's snapshot frozen, which is the right call for a conference record, since a paper list for 2023 does not change after 2023. It also means there is no cross-year view, and no mechanism to see which directions recurred, since each repository starts from an empty table. The curation metadata is thin and worth stating plainly: no GitHub releases, so nothing is versioned as a publication, and the default branch is master rather than main. The last push is dated 15 April 2026, which is recent enough that the list is being touched, and far enough back that a 2023 table may have missed additions made since. The licence is CC0-1.0, a public domain dedication. That covers the curation itself, the table and the notes, and it does not extend to the papers or the code it links, which remain under whatever their authors chose. ## Conclusion Use it as a starting shortlist, not as a survey: nineteen rows against 2,359 acceptances is a taste, and the README itself points at the full accepted-paper list for everything else. Two things to check before you rely on it. The contribution warning tells you to head to CONTRIBUTING.md for how to make changes, and no such file exists at the repository root, which is the one link in the project that does not resolve from the tree. And the entries are preprints, so the arXiv version and the camera-ready differ in places; cite the version you actually read. For a year-by-year survey, the 2024, 2025 and 2026 sibling repositories follow the same format. ## FAQ ### What is SkalskiP/top-cvpr-2023-papers? It is a curated shortlist of CVPR 2023 papers, published under CC0-1.0. Each of the nineteen rows gives a topic, a paper title, a link to the code repository and a link to the paper on arXiv. Sibling repositories hold the equivalent lists for 2024, 2025 and 2026. ### How many papers were accepted at CVPR 2023? CVPR 2023 received 9,155 submissions and accepted 2,359 papers. This repository shortlists nineteen of those, and points readers to the full accepted-paper list on the conference site for anything not included. ### How do I add a paper to top-cvpr-2023-papers? The table is generated and marked as such between AUTOGENERATED_COURSES_TABLE comments, with a warning not to edit it by hand and to consult CONTRIBUTING.md for the procedure. The repository root contains .gitignore, LICENSE, README.md and automation/, and no CONTRIBUTING.md, so the generator script in automation/ is the only guide available. ### Which topics does the CVPR 2023 shortlist cover? The topic column uses labels including Segmentation, Segmentation and Generative AI, NeRF, 3D, Generative AI, Transfer Learning and Benchmarking. They sit at different levels of granularity, and rows of the same topic are not always adjacent, so the column works as a filter rather than as an index. ### Does top-cvpr-2023-papers rank the papers or list award winners? It does neither. There is no rank, score or ordering column, and the selection is the author's shortlist by topic. The repository also does not track conference awards, so award decisions have to be looked up elsewhere. ## Sources - [Issues](https://github.com/SkalskiP/top-cvpr-2023-papers/issues) - [License: CC0-1.0](https://github.com/SkalskiP/top-cvpr-2023-papers/blob/master/LICENSE) - [README](https://github.com/SkalskiP/top-cvpr-2023-papers/blob/master/README.md) - [SkalskiP/top-cvpr-2023-papers on GitHub](https://github.com/SkalskiP/top-cvpr-2023-papers) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/skalskip-top-cvpr-2023-papers