ICCV Camera-Ready

Acceptance at ICCV opens an unusually long runway: decisions landed in late June 2025 and the conference ran October 19–23 in Honolulu, leaving roughly four months between verdict and podium. The camera-ready deadline sits inside that window (the exact 2025 date was not fetchable at verification time — 待核实; it is communicated in the acceptance email and on the conference site). Use the runway deliberately: most camera-ready damage happens in the first rushed week, and most release-promise damage in the last.

The publication pair

An ICCV paper is published through two channels with different jobs:

No article-processing charge exists in this pipeline; costs run through registration. There is also no copy editor: the PDF you upload is what both channels carry forever, so the proofreading pass below is the only one that will ever happen.

De-anonymization is a rewrite pass, not a toggle

Work through these in order, then rebuild and re-read the whole paper:

1. Author block, affiliations, and emails restored in the final class mode of the author kit.
2. Acknowledgements written for real: grants, compute providers, colleagues — everything the double-blind rules banned in March.
3. Self-citations un-contorted where third-person phrasing now misleads.
4. The availability statement upgraded from "code will be made publicly available" (the anonymity-era phrasing ICCV 2025 prescribed) to actual URLs — repository, model weights, project page. This is the moment links become legal.
5. Every rebuttal promise executed. Diff your rebuttal PDF against the final text: each "we will clarify/cite/report X" must correspond to a visible edit, because your reviewers are at the poster session in October.

# Promise-tracking during the camera-ready edit
grep -oiE 'we will [a-z ]+' rebuttal.pdf.txt | sort -u > promises.txt
# ...manually annotate each line with the section where it was honored
awk 'END{print NR" promises to verify"}' promises.txt
pdftotext camera_ready.pdf - | grep -nE '\?\?|Anonymous|under review' | head   # broken refs & leftovers

Numbers that changed since review

Four months is long enough for bug fixes and longer training runs to move your tables. The integrity rule: reviewed numbers may improve in the final version only with a footnote saying what changed and why. Silent upgrades between the version reviewers scored and the version the world reads are how post-publication integrity threads start — at a venue whose proceedings are permanently public on CVF open access, the diff is one script away for anyone.

Synchronizing the public copies

The media embargo ends at acceptance, so the communications sequence inverts:

• Update the arXiv preprint to the camera-ready text once it is frozen; a stale arXiv version outranks the official copy in search for months.
• Choose one canonical BibTeX (the proceedings entry, once it exists) and put it in the repo README; ICCV papers routinely bleed citations across arXiv/ICCV duplicate entries for years.
• Point the project page and repo at the CVF open-access URL when it appears around conference time.
• Any press or lab publicity can now cite an accepted paper — coordinate it with the conference week when attention peaks.

October in another country

ICCV's venues rotate internationally (Honolulu 2025; Hong Kong announced for October 2027), which makes presentation logistics part of camera-ready season, not conference week:

• Registration: confirm the current cycle's rules on paper registration and in-person presentation — these were not verifiable for 2025 (待核实) and vary across CVF venues; the acceptance email is authoritative.
• Visas: start the moment the decision arrives; some consulate queues are longer than the June-to-October gap (planning ledger in iccv-workflow).
• Tier deliverables: an oral needs a timed talk plus poster; a Highlight poster draws extra traffic and must narrate the paper alone; every poster faces a multi-thousand-person hall where one legible headline result beats four dense panels.
• Backup presenter: name one when travel is uncertain; a no-show poster helps nobody and can breach presentation obligations.

Final-version defect list

Recurring faults in published CVF-venue PDFs, all catchable in an afternoon: review-mode line numbers surviving into the final build; \usepackage[review] left on; figures downsampled by a compression pass until the qualitative comparison is invisible; references still citing arXiv identifiers for papers long since published (fix per iccv-related-work); author-order or affiliation mismatches between the PDF, OpenReview, and the publisher form — the mismatch propagates into IEEE Xplore metadata and is painful to correct after indexing.

Reverify for the current cycle (all 待核实 for 2027)

• Camera-ready deadline, page allowance for the final version, and whether an extra page is granted for acknowledgements.
• IEEE copyright form mechanics and PDF validation requirements.
• Registration-per-paper and presentation-mode rules.
• Poster dimensions, talk lengths, and any virtual-platform uploads.

Output format

[Camera-ready state] runway weeks left: <n>; blockers: <list>
[Promise diff] rebuttal commitments honored: n/m
[Number changes] none / footnoted: <table → reason>
[Links live] repo · weights · project page · availability statement updated
[Public sync] arXiv refreshed · canonical BibTeX set · CVF URL linked
[October plan] registration · visa · tier deliverables · backup presenter