SkalskiP/top-cvpr-2025-papers: a curated CVPR 2025 reading list with paper, code and demo links
About This repository is a curated collection of the most exciting and influential CVPR 2025 papers. 🔥 [Paper + Code + Demo]
At a glance
- What is it?
- The repository is a hand-picked index of CVPR 2025 papers grouped by topic, with each entry linking to arXiv, a code repository, a video and sometimes a demo. It is a discovery aid, not a paper database, and its value depends entirely on how much you trust the shortlist.
- Who is it for?
- Adopt it as a starting shortlist if you want a small, opinionated slice of CVPR 2025 with code and demo links already attached, and treat the topic headings as the entry points. Do not use it as a substitute for the official accepted-papers list if you need completeness, and do not quote it as a ranking of paper quality, because the README does not claim one.
- Can I use it commercially?
- Yes. CC0-1.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 158 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 20, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the shortlist is actually for
CVPR is large enough that the accepted-paper list is not a reading list. The README states that 13,008 papers were submitted for 2025 and 2,878 were accepted, and the author says the repository exists to help you search for the best of those publications. That framing matters: the project is a filter, not an archive. Each entry is a person's judgement about which papers are worth your attention, and the README points readers whose paper is missing to the full accepted-papers list on cvpr.thecvf.com.
The audience is narrow and identifiable. You are a computer vision engineer or researcher who wants to see what shipped at CVPR 2025 in areas such as 3D vision, depth estimation, image segmentation, object detection and vision-language models, and you want the code and demo links next to the paper link rather than hunting for them. If you need every accepted paper, or you need the official program metadata, this repository is the wrong starting point, and the README says so itself.
How the entries are structured and generated
The visible unit is a block per paper. It contains a poster image, the title with a flame marker for highlighted papers, the author list, a row of bracketed links (paper, code, video, and demo where one exists), a Topic line, and a Session line with the day, time, timezone and poster number. For example, the VGGT entry lists a paper link to arXiv, a code link to facebookresearch/vggt, a video link and a Hugging Face demo, with the topic 3D Vision and a Friday poster session slot.
The README carries an explicit warning inside the generated list: the list is autogenerated and should not be edited manually, with a pointer to a CONTRIBUTING.md file in the repository for how to make changes properly. The repository layout shows an automation/ directory at the top level alongside CONTRIBUTING.md and LICENSE. So the papers list is produced by tooling in this repository, and contributions are expected to go through that path rather than by editing README.md directly. The primary language is Python, which is consistent with the automation directory being the place where generation happens. What the README does not document is the schema of the source data the generator reads, so anyone wanting to change the selection criteria has to read the automation code rather than a spec.
Reading the list without installing anything
There is no package to install and no service to run. The README gives no installation steps because the artifact is the README itself, and the only repository files it points contributors at are CONTRIBUTING.md and the automation/ directory. The practical first use is to open the README on GitHub and read the topic headings, which are the only navigation the project provides.
The headings visible in the README include 3d vision and depth estimation, each followed by paper blocks. Within a block, the bracketed link row is the part you act on: the paper anchor goes to arXiv, the code anchor goes to a GitHub repository, and where present the video and demo anchors go to YouTube and Hugging Face Spaces. The VGGT block, for instance, carries all four. If you want to work on the list itself rather than read it, the README's own instruction is to head to CONTRIBUTING.md for details on how to make changes properly; it does not document a generator entry point, so that file is the starting point rather than a command.
Where the curation model breaks down
The list is a snapshot of one person's attention, and the README does not describe a selection rubric. There is no stated criterion for what makes the shortlist, no note on how many papers were considered, and no changelog of what was added or removed. The flame marker for highlighted papers is likewise unexplained in the README. That means you cannot audit the filter, only accept or reject it.
The topic taxonomy is coarse. Headings like 3d vision and depth estimation sit next to each other, and a paper that straddles both is filed under one. A reader looking for a specific subfield has to scan blocks rather than query a field. There is also no structured data file documented in the README; the entries live as HTML inside Markdown, which is why the block for each paper is a paragraph element with an image and a set of anchors. If you wanted to build a searchable index over this list, you would be parsing generated HTML in Markdown, not reading a JSON or CSV that the project publishes.
Finally, the session lines are conference metadata. They are useful for finding a poster at the venue, and useless once the conference is over. Nothing in the README says these lines are refreshed or pruned.
Compared with the official accepted-papers list
The real alternative is the source the README itself names: the CVPR 2025 accepted-papers page on cvpr.thecvf.com. The difference is one of purpose, not quality. The official list is exhaustive and authoritative; it tells you what was accepted and nothing about whether a paper has code, a demo or a video. This repository is the inverse: a small selection where the author has already attached those links, at the cost of completeness and of any stated selection rule.
A second, less direct alternative is the sibling repositories the README links in its header: the 2023, 2024, 2025 and 2026 editions of the same project. They share the format and the topic headings, so if you are tracking a research area across years, moving between them is cheap. What they do not give you is a merged view; each year is a separate README with its own generated list.
If your need is exhaustive coverage with structured metadata, neither this repository nor the official HTML page is a great fit, and you should look at the conference's own program data instead. If your need is a short list with code attached, the official page cannot help you at all.
Licence, maintenance and what a fork costs you
The repository is licensed CC0-1.0. For a curated list of links and titles, that is the permissive end of the spectrum: the licence removes copyright restrictions on the repository's own content. It does not, and cannot, cover the papers, images, code repositories or demos that the entries link to. Those remain under their own terms, and the poster images in particular are hosted on external storage, not in this repository. Reusing the list is therefore straightforward; reusing anything it points at is a separate question for each linked project.
The last push to the repository was on 2026-04-15. That is roughly five months before today, so the repository is not archived, but the README gives no release history and no versioning scheme, so there is nothing to upgrade in the usual sense. The cost of adopting it is the cost of tracking a moving README: if you fork it, you inherit the generated list and the automation directory, and you own the merge conflicts when the upstream list changes. If you only read it, the cost is the time you spend checking that a linked code repository still exists, since the README does not state that links are validated.
Editorial conclusion
Adopt it as a starting shortlist if you want a small, opinionated slice of CVPR 2025 with code and demo links already attached, and treat the topic headings as the entry points. Do not use it as a substitute for the official accepted-papers list if you need completeness, and do not quote it as a ranking of paper quality, because the README does not claim one. Before relying on any entry, open the linked arXiv page, confirm the code repository exists and is reachable, and check the session line to see which day and poster number the paper belongs to.
Frequently asked questions
What are some famous CVPR papers?
The repository does not rank papers by fame or citation. It marks some entries with a flame symbol as highlighted papers, but the README does not explain what that marker means, so it should be read as the author's emphasis rather than a measure of influence.
Is CVPR considered a top conference in computer vision?
The README frames CVPR as a massive conference, stating that 13,008 papers were submitted for 2025 and 2,878 were accepted. That submission volume is the only signal the material gives about its standing, and the repository does not make a ranking claim.
What is the CVPR acceptance rate?
The README reports 13,008 submissions and 2,878 accepted papers for 2025. It does not state an acceptance rate, so you would have to compute it from those two figures yourself.
Community notes