How to control duplicate files without deleting useful variants

This scenario does not demonstrate a 90% reduction. It explains how to review similar files without assuming that everything alike is a duplicate.

A repository contains “acme_logo.png”, “acme_logo_copy.png”, “acme_logo_white.png”, “acme_logo_2022.png”, and “acme_logo_FINAL.ai”.

The first two may be exact copies. White is a variant. The 2022 file may be obsolete. AI may be the primary editable. Deleting four files by visual similarity would destroy useful context.

Classify the relationship

• Exact duplicate: binary-identical content under any filename.
• Different compression: visual content matches but resolution, quality, or metadata differs.
• Version: another stage such as draft, review, approved, or published.
• Adaptation: a different format, language, channel, size, or territory linked to a master.
• Similar but independent: related photographs or designs that are not duplicates.

Do not delete automatically

Detection can create a review queue: group candidates, show metadata and use, identify related campaigns, suggest a primary asset, let a person decide, and record the decision.

Regulated, published, or differently licensed files require more care.

Choose a source of truth

For each family, identify the editable master, approved version, linked adaptations, originals, and obsolete versions.

Redirect references to the primary asset before removing a copy. Do not leave presentations or campaigns pointing at a deleted file.

A photograph example

Four visually identical JPGs appear: 6000 × 4000 master, 2000 × 1333 web, 1080 × 1350 Instagram, and 1080 × 1350 advertisement with text.

They are one original and three derivatives, not four useless copies. Record the relationship so nobody uses the advertisement as a clean photograph.

Establish a baseline

Measure confirmed exact copies, families without a primary asset, versions with unknown state, detached adaptations, recovered space, and obsolete-file incidents.

Improvement is the difference between reviews, not a universal figure.

Similarity in an asset library helps discover candidates. Governance decides what each relationship means and what can be removed.