top-cvpr-2025-papers is a generated shortlist, and the warning inside it sends you to last year's contributing guide

The do not edit warning points at a different year's repository

The list section opens with a marker comment and a warning block. The marker names the region as automatically generated, and the warning underneath says the list must not be edited by hand because it is produced automatically, and then it tells you to head over to a contributing guide hosted in the 2024 repository for details on how to make changes properly.

That is the wrong repository. This one has its own contributing file at the root. The link points sideways into a sibling repository one year earlier, presumably because the four repositories share a template and the link was written once and copied forward without updating the year.

The practical effect is small but annoying: following the project's own instruction lands you in documentation for a different year's list. It also tells you something about the structure, since the four sibling repositories are the header links for 2023, 2024, 2025 and 2026, and they evidently share both the entry format and the template.

Thirteen thousand submissions and a short list on top

The opening paragraph gives the two numbers that frame everything else. In 2025 alone, 13,008 papers were submitted and 2,878 were accepted, which by arithmetic is a little over a fifth. The author then states the purpose plainly: to help search for the best of the CVPR publications.

The word used is a short list, and the same paragraph says what to do when a paper is not on it, which is to take a look at the full list of accepted papers on the conference site. That is an unusually clean framing for a curated repository, because it establishes in the first paragraph that this is a subset with no claim of completeness and that the authoritative source is one click away.

It also sets the expectation correctly. A repository of this kind is usually read as a ranking, and the numbers here show why that reading fails: even a very selective short list against 2,878 accepted papers is going to leave out work that matters to a given reader. The value is in the topic grouping and the outbound links, not in the omission.

A highlight mark lands on about two entries in five

The list uses one mark for curation and defines it above the entries. A highlighted marker is explained as highlighted papers, and it is applied by wrapping the entry title in bold and prefixing it with the same mark in the list.

In the sections that are visible, the rate is consistent. The 3D vision section has three entries and two of them are marked, a geometric transformer and a dense simultaneous localisation and mapping system with reconstruction priors, while a radiance field relation paper is not. The depth estimation section also has three entries with two marked, a long consistent depth sequence generator and one interpretation paper, while a universal camera paper and a super long video depth paper are not.

So roughly two in five, by the visible sample. No criteria for the mark are written anywhere, which is the honest limitation: a reader cannot tell whether it tracks oral status, awards, citations, or the author's own taste. It is a signal that the author found something worth marking, and nothing more.

Every entry carries its poster session slot and a PDT hour range

Each entry ends with two lines that are not links. One is the topic, which is the section heading the entry sits under, and the other is the session, which gives a weekday, a date in June, a start and end time in Pacific time, a poster session number, and a position within that session.

The numbers in the visible entries span five separate poster sessions. The marked geometric transformer is in the second session, the marked simultaneous localisation system in the fourth, a radiance field paper in the fifth, and the depth entries split across the first and the fifth. Two sessions on the Friday, one on the Saturday, and two on the Sunday. Times are given in half hour to two hour blocks, with the earliest visible start at half past eight in the morning.

That is unusually granular for an index. It means you can work out which day to attend if you only care about one topic, and it is the one piece of information in the file that is not a link.

Nine lines of HTML per paper, and the demo link is optional

The entry format is a fixed HTML block, and reading one entry tells you the schema for all of them. It opens with an anchor pointing at a conference-hosted poster PDF thumbnail, with a cache-busting timestamp appended as a query parameter. Then comes an anchor on the paper's preprint identifier carrying the title as its tooltip. Then a line of authors. Then a line of up to four labelled links, for the paper, the code, a video, and a hosted demo. Then the topic and the session lines.

The four link types are not all present on every entry, and the pattern is worth knowing so you do not read a missing link as a broken one. The simultaneous localisation entry has a paper, code and video but no demo. The radiance field entry likewise stops at video. The universal camera entry and both of the depth entries with demos have a paper, code and a hosted demo space but no video. So a paper can have a demo and no video, or a video and no demo.

Because the whole block is HTML rather than markdown, the rendered page is a wall of markup with no table of contents. Searching for a paper name in the rendered view, or reading it in a terminal without markup rendering, is the practical way to use it.

Four files and one directory, and Python is the generator

The repository is small on purpose. The top level holds a readme, a contributing file, a licence, a gitignore, and one directory called automation. That directory is why the recorded primary language is Python: the published artefact is the readme, and the Python is the thing that writes it.

That arrangement has an obvious consequence. If a poster session time is wrong, or a paper is missing a code link, or the highlight mark is wrong, editing the readme means fighting an automated process. The warning comment says as much. It also means the repository has no tests and no build to speak of, and that the quality of the output depends entirely on whatever the generator scrapes from the conference site.

The licence is CC0, the public domain dedication rather than a conventional permissive licence. For a list of links and paper titles that is a defensible choice, since the page is mostly metadata pointing at other people's work, and it means anyone can fork the year-specific repository and re-point the generator without attribution obligations.

Four sibling repositories, and this one is neither first nor last

The header is a single row of four links, one per year, spanning 2023 through 2026. The naming is identical apart from the year, and the 2026 entry exists, which means this 2025 list is neither the first of the series nor the current one. It is the third.

That matters for anyone citing or reusing the pattern. The generator presumably runs the same way for all four, so the 2026 repository is the one to watch if you want to see how the automation ages, and the 2023 one is the one to read if you want to see how the entry format evolved. All four links in the header, and the contributing link inside the warning, point at the master branch, so the default branch name is consistent across the series.

There is no version history to speak of: the repository publishes no releases, the default branch is master rather than main, and the last commit landed on 2026-04-15, which is months after the conference the list covers.

Editorial conclusion

Use it as a shortcut into a topic rather than as a ranking. The selection is a personal short list, the author says so, and the full accepted list is one link away for anything missing. Two things to know before you rely on it. The list is machine generated, so correcting an entry means finding the generator rather than editing the readme, and the in-file warning about that points at a different year's repository. And the highlight marks are a curation signal, not a ranking: the author applies them to a subset, and no criteria for that subset are written down, so treat a marked paper as a suggestion rather than as an award.

Frequently asked questions

Official sources

1. Issues
2. License: CC0-1.0
3. README
4. SkalskiP/top-cvpr-2025-papers on GitHub