Define an editing shortlist, not client approval
This workflow is for photographers and event teams selecting source material for editing. Its output is a human-approved shortlist, a selection report and an identified file package, not a finished album or a client proof PDF. Agree who receives it, which moments matter and whether alternative expressions or intentional motion blur should remain. Back up originals before starting; exclusions here are decisions, not permission to delete.
Inventory format, size and metadata and assign a batch ID plus source path and SHA-256. Identical camera basenames are not unique identities. The culler accepts JPEG, PNG and WebP up to 100 files, 20 MB each, 100 MB total and 40 MP decoded per image. Split larger sets deliberately and record group coverage: similar photos in different batches are not compared together. RAW development belongs upstream; map each preview back to its original.
Check orientation and privacy before analysis
The culler processes photos locally without uploads or AI. Supporting tools may operate on a server, so obtain permission before sending private photographs or reports containing filenames. Inspect orientation as actually displayed. If a correction is needed, auto-orient a working copy and compare the result; record its path and hash. The shortlist then points to that prepared input, not byte-identically to an untouched camera original.
Do not remove metadata merely to make orientation appear consistent. Resolve orientation first and preserve source identity. The selected ZIP does not perform redaction: it retains input bytes, including metadata. Review faces, names, locations, EXIF and filenames separately. If the editor may not receive those originals, create privacy-approved derivatives through a separate metadata workflow, verify them and use new hashes. Keep restricted originals outside the handoff.
Use scores to find questions, not answers
The tool measures heuristic signals from a 256-pixel analysis thumbnail. Edge detail is a relative multi-scale signal rather than the probability that the subject is in focus. Dark and bright fractions count thumbnail luminance at or below 8 and at or above 247; they do not measure recoverable RAW highlights, calibrated exposure or full-resolution clipping. Noise, bokeh and deliberately bright or dark scenes can skew the signals.
Review the important subject at full size and compare similar frames. The grouping is conservative near-duplicate detection, not recognition of every photograph from one scene. It has no face or closed-eye detection. Two expressions may deserve separate keeps; visually similar crops may have different uses. The initial state keeps a recommended member of each group, so explicitly inspect and override those defaults before exporting. Recheck choices after regrouping or reanalysis rather than assuming earlier selections persist.
Record keep, exclude or unresolved decisions with reasons in a separate ledger. The exported report contains recommendations and selection flags, not your complete editorial rationale. Investigate decode failures individually; they are not evidence of poor photographic quality. Preserve originals and useful alternatives even when the shortlist is small.
Reconcile the CSV before trusting the package
Download the report after the last decision. Validate CSV quoting and columns, then reconcile the original batch count with analyzed and failed rows. Among analyzed rows, confirm kept and excluded counts. Check filename, group, kept, recommended, dimensions and sha256 against the batch ledger. Failed rows have empty metric fields and an explanation; do not silently treat them as normal rejections. Group numbers are local to a run, not durable cross-batch identities.
Use batch plus hash to resolve repeated filenames. Preserve the report as generated and keep human reasons alongside it. A structurally valid CSV says nothing about expressions, composition or completeness of the ZIP. Download the ZIP from the same final selection; label both files with a shared batch/revision name and retain all reports when splitting a shoot into batches.
Restore and identify the delivered photos
Preview the ZIP and compare its entries with the approved kept set. Photo entry names have a numeric prefix and a sanitized original name; truncation and renaming mean filename equality alone is insufficient. The archive also contains photo-report.csv, which can include excluded and failed inputs. Check whether that report itself exposes information the recipient should not receive.
Extract selected entries by exact archive path for inspection, then reconcile every delivered photo, using local extraction when convenient. Open the files and compare SHA-256 to their intended inputs; the byte-preserving export should match. A few spot checks do not prove the entire package. If a photo was deliberately transformed for delivery, verify against its approved derivative instead and document the link to the original.
Hash the final ZIP and report after all changes. Include reviewer, decisions, failures, privacy status and unresolved cases; temporary processing links are not durable delivery. Continue with RAW preparation for source development, metadata review for privacy, or the photo contact-sheet workflow for client selection. This workflow stops at the editor's verified candidate set, not retouching or final artistic approval.