Guides · Free tool
Drop your photos. See what a recall-first cull looks like.
50 to 500 JPEGs, straight from a shoot. This page groups your bursts, estimates your keeper ratio, and points at the frames a first pass would question — using deterministic image math, right here in your browser.
Nothing leaves your browser. Structurally.
There is no upload in this tool — no server receives your images, full stop. Don't take our word for it: load this page, switch your machine to airplane mode, then drop your photos. Everything still runs. (Site analytics counts visits and photo counts by range — never image data. Block it and the tool works exactly the same.)
Drop your JPEGs here
or
JPEG only in this version (that's what a culling pass starts from). RAW and HEIC files are skipped. Photos are read locally, analyzed at 800 px, then released from memory.
Your audit
–
photos analyzed
–
burst groups (2+ near-identical frames)
–
keeper candidates (best of each burst + clean singles)
–
estimated keeper ratio
Where that ratio sits, against one documented wedding
The only fully published reference we quote: in a December 2024 review (mariess.co.uk), a wedding photographer's own final pick was 580 keepers out of 4,800 frames — about 12%. The AI culling pass they reviewed kept 1,500 of the same 4,800 — about 31%, nearly three times their real selection. One documented wedding, not a statistic; it's still the clearest published number on what over-selection looks like.
What one manual pass over this set costs you
Plain arithmetic, not a statistic: – photos × 2.0 s per decision ≈ – for a single review pass. Adjust the pace to yours:
Your burst groups
Frames shot within 2 seconds of each other that also look near-identical. The highlighted frame is the sharpest of each group — one keeper candidate per burst, however many frames the burst holds.
Frames a first pass would question
Softest frames of the set (relative to your own set's median) and exposure outliers (blown highlights / crushed shadows). A flag is a reason to look, never a verdict — the frame where the moment happened can absolutely be worth a little blur.
This audit reads optics. BestTake reads the moment.
Everything above is deterministic math — sharpness, exposure, near-duplicates. What it can't see is which frame holds the first kiss. That's the actual job: a recall-first cull that surfaces the moment, runs on your own machine, and hands your picks back as XMP.
How it works — the exact method
No black box: this tool runs four classic, deterministic image measures, and you can hold it to them.
- Sharpness — variance of the Laplacian. Each photo is decoded at 800 px and convolved with the 4-neighbour Laplacian kernel; the variance of the response is the standard blur estimator from the computer-vision literature. Higher = more edge energy = sharper. Scores are compared within your set (a frame is "soft" below 30% of your set's median), because absolute sharpness depends on lens and subject.
- Near-duplicates — difference hash (dHash). Each photo is reduced to a 9×8 grayscale grid; each cell is compared to its right neighbour, giving a 64-bit fingerprint. Two frames are "near-identical" when their fingerprints differ by ≤12 bits (≤6 when no capture time is available).
- Bursts — EXIF capture time. Frames taken within 2 seconds of each other and near-identical by dHash are grouped as one burst. Capture time is read from the EXIF DateTimeOriginal tag in your files, locally. Files without EXIF fall back to their file-modified time.
- Exposure — histogram tails. The share of pixels at gray level ≥250 (blown) and ≤5 (crushed). A frame is flagged past 15% blown or 40% crushed.
- Keeper estimate. One keeper candidate per burst (its sharpest frame) plus every unflagged single frame. That's a floor estimate of a decisive first pass — your taste will rightly overrule it in both directions.
What this tool deliberately doesn't do: no machine learning, no face or closed-eye detection, no "moment" recognition. Those need models, not arithmetic — closed-eye detection is on our list for a future version of this free audit. Keeping the audit deterministic is what lets it run instantly, offline, in a tab.
Reference numbers: one documented wedding from a published December 2024 review (mariess.co.uk) — 4,800 frames, 580-image final pick, 1,500 kept by the reviewed AI pass. Quoted as a single dated testimony, not a statistic. Time math is plain multiplication at a pace you set yourself.
Further reading
Estimates from deterministic image measures on your own files — your judgment decides. Nothing on this page is stored or sent anywhere; analytics counts visits and photo counts by range only. Content written with AI assistance, reviewed before publication by the publisher.