cvpr-writing-style
GitHub指导CVPR论文写作与修订，涵盖8页含图限制、首图惯例、基准表制作及针对怀疑主义审稿人的声明校准，强调视觉优先的两遍阅读体验及工具辅助责任。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill cvpr-writing-style -g -y
SKILL.md
Frontmatter
{
"name": "cvpr-writing-style",
"description": "Use when writing or revising a CVPR paper's prose and figures, covering the 8-page budget where figures and tables count against the limit, the page-1 teaser figure convention, benchmark-table craft, claim calibration for a skeptical vision audience, and taking responsibility for tool-assisted text under CVPR's rules."
}
CVPR Writing Style
CVPR is a visual venue with a brutal budget: eight pages including figures and tables (references overflow free — verified 2026 rule). The papers that survive triage by an overloaded reviewer pool are the ones a reader can evaluate in two passes: a 90-second flip through figures and tables, then a real read. Write both passes on purpose.
The two-pass reader
Pass one is visual. Reviewers at a 16k-submission venue flip: Figure 1, the tables, the qualitative grid, the ablation. If that flip communicates problem → idea → win, the close read starts friendly. Concretely:
- Figure 1 (the teaser) sits on page 1 and must state the paper alone: task, what prior work gets wrong, what yours does instead — ideally on the same input. A teaser that needs the caption's third sentence has failed.
- Every caption is self-contained: dataset, metric, and the takeaway sentence. Flip readers never see body text.
- The main table is typeset as an argument: best numbers bold, second-best underlined, methods grouped by supervision/backbone class, your row where the eye lands last.
Budgeting eight inclusive pages
| Section | Typical share | Compression lever |
|---|---|---|
| Title/abstract/teaser | ~1 page | Teaser earns its half page or shrinks |
| Intro | ~1 page | Contributions as 3 bullets, no history lesson |
| Related work | ~0.75 page | Position, don't survey (see cvpr-related-work) |
| Method + pipeline figure | ~2 pages | One figure replacing a page of prose |
| Experiments | ~2.5–3 pages | Grids to supplement; keep the decisive ablation |
| Limitations/conclusion | ~0.25–0.5 page | Honest and short |
Because figures count, figure design is text editing: a two-column pipeline figure costing 0.6 page must delete at least 0.6 page of method prose, or it is negative space. Cut from the flip-reader's perspective — never cut the ablation table to save a paragraph of adjectives.
Sentence-level norms of the genre
- Present tense, active voice, mechanism first: "We replace the cost volume with …" beats "In recent years, cost volumes have been widely used…".
- Name the delta in numbers early: "+2.1 AP on COCO with 40% fewer FLOPs" in the abstract, not "significant improvements".
- One notation table, introduced once; vision papers drown in re-defined symbols.
- Kill hype adjectives ("novel", "remarkable", "extensive") — at this venue they mark inexperience; the table is the adjective.
Claim calibration
Vision reviewers are professionally allergic to overclaiming, because the field's benchmark culture makes checking easy:
Weak claim, strong paper:
"Ours improves X on datasets A and B under protocol P with backbone Q."
Strong claim, dead paper:
"Ours solves X" / "generalizes to real-world scenes" (shown on 2 curated clips)
"state-of-the-art" (on one dataset, vs. year-old baselines)
"real-time" (no hardware named — CRF row contradicts you)
Every superlative must name its scope: dataset, protocol, compute class. If a limitation is visible in your own qualitative grid, write it in the limitations paragraph first — reviewers reward the paper that scoops its own weaknesses.
Tool-assisted text is your text
The verified 2026 author-side rule: use whatever tools you like, but fabricated citations and factual inaccuracies are rejection grounds — possibly without review — and "an LLM did it" is explicitly no defense. Operational habit: any machine-drafted paragraph gets a human pass for (a) citations that exist and say what's claimed, (b) numbers matching your own tables, (c) generic filler that flip-readers instantly recognize as padding.
An abstract skeleton that fits the venue
Six sentences carry a CVPR abstract; more usually means the contribution is unclear:
- Task + failure: the specific visual problem and where current methods break ("Trackers lose identity during occlusions longer than their temporal horizon").
- Cause: the mechanism behind the failure, in one clause — this sentence is what separates insight papers from tweak papers.
- Method: what you built, named, with its one central idea.
- Headline evidence: the scoped number ("+2.1 AP on COCO, matched backbone").
- Breadth or cost: second dataset, transfer result, or the efficiency figure.
- Availability: what will be released (code/models/data — and remember datasets claimed as contributions carry a camera-ready release obligation).
Write the abstract twice: once before the experiments as a plan, once after as a report; the diff between the two is your overclaiming audit.
Revision protocol for a drafted paper
- Print the figures and tables alone; run the 90-second flip test on a colleague outside the project.
- Reconcile abstract numbers against tables (rebuttal-week embarrassment prevention).
- Hunt "see supplement" pointers that hide decision-critical logic; pull those back.
- Delete the roadmap paragraph ("Section 2 reviews…") — page-limited venues stopped reading them years ago.
- Read the intro's last bullet against the conclusion: the promised contribution and the delivered one must be the same sentence.
Reverify each cycle
- Page cap and the references-only overflow rule.
- Template/author-kit version (build from the current
cvpr-org/author-kit). - Any new writing-adjacent form fields (limitations sections, AI-use disclosures — none verified for 2026, 待核实 for later cycles).
Output format
[Flip test] passes / fails at: <figure or table>
[Budget] pages by section vs 8.0 incl. figures
[Claim audit] superlatives without scope: <list>
[Caption audit] non-self-contained: <figure numbers>
[Tool-text risk] unverified citations or numbers: <locations>
[Edit order] <highest-leverage cuts first>
Version History
- 9f86f09 Current 2026-07-19 14:42
