Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when preparing the final version of an accepted CVPR paper, covering de-anonymization, the dataset-release-by-camera-ready obligation, CVF open access versus the IEEE Xplore version of record, oral/highlight/poster preparation at conference scale, and arXiv/preprint synchronization after acceptance.
.claude/skills/brycewang-stanford-cvpr-camera-ready/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 79% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 4% | 0% |
Use this after a CVPR acceptance (decisions landed February 20 in the 2026 cycle, conference in early June). The verified 2026 publication pipeline is described below; the mechanical camera-ready details for any given year — exact deadline, page allowance, IEEE forms — were not fetchable at check time and are 待核实: read the instructions email and the current Author Guidelines before building the final PDF.
Accepted CVPR papers are published twice, and the two copies serve different audiences:
openaccess.thecvf.com): free to everyone, posted around theconference; the file is the accepted version apart from a watermark. This is the copy the field reads and the one bibliography tools usually find first.
Institutional reporting and some citation databases key off this one.
There is no author-side publication fee; costs ride on registration. Consequence for you: the PDF you upload at camera-ready is the artifact of record in both channels — there is no copy-editing pass to catch your mistakes.
Reversing double-blind is edit work, not a switch:
now.
showed" → "we showed in 12]" only where it clarifies lineage).
page, dataset host. Links were banned at submission; the camera-ready is where they belong.
tight page over whatever the final-version limit is that year.
Verified 2026 policy: if the paper claims a dataset release as a scientific contribution, the dataset must be publicly available no later than the camera-ready deadline. "Available upon request" or a landing page with a form does not meet the bar you promised reviewers. Licensing, hosting bandwidth, and consent/PII scrubbing all have to be finished on the paper's schedule, not the lab's.
bash# Release-readiness probe, run from a machine outside your institution curl -sIL https://<dataset-host>/<archive> | grep -Ei 'HTTP|content-length' sha256sum downloaded_archive.tar.gz # matches the checksum printed in the repo? python3 - <<'EOF' import json; m = json.load(open("meta/license.json")) assert m["license"] and m["contact"], "license/contact missing from release metadata" EOF
In 2026 the program split acceptances into tiers (program trackers reported 141 orals in four parallel tracks and 578 highlights among 4,090 acceptances — cross-check the official program before quoting). What each tier means for preparation:
| Outcome | Extra deliverables | Preparation reality | |---|---|---| | Oral | Talk + poster | A few minutes on a big stage; rehearse to time, design slides for the back row | | Highlight | Poster, flagged in the program | Extra foot traffic; the poster must carry the paper alone | | Poster | Poster | Thousands of attendees walk the halls; a legible headline result beats density | | Award candidate | Talk logistics vary | Committee re-reads the paper — camera-ready polish matters twice |
Every tier presents at a venue with five-digit attendance (about 12,200 registrants in 2026), so print deadlines, poster-size specs, and the video/teaser uploads requested by the virtual platform belong on the same checklist as the PDF.
The version of record gets no editorial pass, so run your own. Faults found in published CVPR PDFs year after year, all preventable in an afternoon:
will cite Foo]", "we will clarify the protocol") against the final text, item by item.
(bug fix, longer training), say so in a footnote; unexplained deltas between the reviewed and published versions are the stuff of integrity threads.
\usepackage{review}-modeartifacts from the author kit.
Fig. ??), and citations stillpointing to arXiv versions of papers that have since been published (fix per cvpr-related-work).
file with the publisher must agree on names, order, and affiliations — mismatches propagate to IEEE Xplore and are painful to correct later.
After acceptance, converge the copies: update the arXiv version to match the camera-ready (arXiv posting was already legal during review), point the repo README at the CVF page once it exists, and use one canonical BibTeX everywhere — mixed arXiv/proceedings citations fragment your citation count for years. Keep the media embargo in mind in reverse: it lifted at acceptance, so coordinated publicity is now allowed.
copyright block on).
text[Camera-ready status] on-track / blocked [Version-of-record checks] deanonymized · acknowledgements · live links · page fit [Dataset clause] not-claimed / released-and-probed / BLOCKED: <gap> [Tier prep] oral/highlight/poster deliverables and dates [Public sync] arXiv updated · repo pointed · canonical BibTeX chosen [待核实 items] <forms, deadline, page allowance for this year>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-07 | pass→pass | 22,362 | 20,904 | -7% | 1 | 1 | 0% | 2,543 | 3,738 | +47% | 0 | 0 | — |
case-01 | fail→pass | 22,084 | 28,869 | +31% | 1 | 1 | 0% | 2,734 | 4,900 | +79% | 0 | 0 | — |
case-02 | fail→fail | 17,671 | 17,895 | +1% | 1 | 1 | 0% | 1,756 | 3,220 | +83% | 0 | 0 | — |
case-03 | fail→fail | 27,848 | 20,079 | -28% | 1 | 1 | 0% | 3,573 | 3,495 | -2% | 0 | 0 | — |
case-04 | fail→pass | 19,872 | 16,686 | -16% | 1 | 1 | 0% | 2,810 | 3,233 | +15% | 0 | 0 | — |
case-05 | fail→pass | 24,772 | 16,632 | -33% | 1 | 1 | 0% | 3,955 | 4,010 | +1% | 0 | 0 | — |
case-06 | pass→pass | 19,866 | 15,030 | -24% | 1 | 1 | 0% | 2,349 | 3,096 | +32% | 0 | 0 | — |
case-08 | pass→pass | 18,350 | 13,927 | -24% | 1 | 1 | 0% | 1,996 | 3,512 | +76% | 0 | 0 | — |
case-09 | pass→pass | 23,280 | 12,061 | -48% | 1 | 1 | 0% | 2,799 | 3,050 | +9% | 0 | 0 | — |
case-10 | fail→pass | 18,562 | 16,508 | -11% | 1 | 1 | 0% | 1,958 | 2,956 | +51% | 0 | 0 | — |
case-11 | pass→pass | 18,090 | 12,647 | -30% | 1 | 1 | 0% | 1,810 | 2,587 | +43% | 0 | 0 | — |
case-12 | fail→fail | 22,992 | 21,359 | -7% | 1 | 1 | 0% | 2,384 | 3,640 | +53% | 0 | 0 | — |
case-13 | pass→pass | 22,364 | 25,389 | +14% | 1 | 1 | 0% | 2,667 | 3,348 | +26% | 0 | 0 | — |
case-14 | pass→pass | 21,146 | 19,418 | -8% | 1 | 1 | 0% | 2,980 | 3,537 | +19% | 0 | 0 | — |
case-15 | fail→fail | 30,725 | 12,527 | -59% | 1 | 1 | 0% | 3,698 | 2,966 | -20% | 0 | 0 | — |
case-16 | fail→pass | 25,467 | 15,608 | -39% | 1 | 1 | 0% | 2,985 | 3,119 | +4% | 0 | 0 | — |
case-17 | pass→pass | 31,176 | 20,261 | -35% | 1 | 1 | 0% | 3,096 | 3,436 | +11% | 0 | 0 | — |
case-18 | pass→pass | 36,774 | 26,836 | -27% | 1 | 1 | 0% | 4,013 | 4,268 | +6% | 0 | 0 | — |
case-19 | pass→pass | 19,852 | 17,063 | -14% | 1 | 1 | 0% | 2,291 | 3,376 | +47% | 0 | 0 | — |
case-20 | pass→pass | 25,810 | 26,608 | +3% | 1 | 1 | 0% | 3,111 | 4,565 | +47% | 0 | 0 | — |
case-21 | pass→pass | 26,134 | 27,361 | +5% | 1 | 1 | 0% | 2,521 | 4,429 | +76% | 0 | 0 | — |
case-22 | pass→pass | 25,164 | 20,078 | -20% | 1 | 1 | 0% | 2,583 | 3,661 | +42% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +23 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.