Remove Similar Image

Remove Similar Image

Core Goal

• Scan one local image or a directory tree of local images.
• Use ImageHash to cluster exact duplicates and near-duplicates.
• Use OpenCV variance-of-Laplacian blur scoring to flag blurry shots.
• Preview actions first, then permanently delete or move candidates to a trash folder when requested.

Required Script

• Use scripts/remove_similar_images.py.
• Start with doctor if dependency availability is unknown.
• Treat analyze without --apply as the safe default.
• Prefer --trash-dir before permanent deletion so the user can review results.

Dependency

• This skill requires Pillow, ImageHash, numpy, and OpenCV:

python3 -m pip install Pillow ImageHash numpy opencv-python-headless

• If doctor reports missing dependencies, stop and surface the install command instead of pretending the scan ran.

Workflow

1. Check dependencies:

python3 scripts/remove_similar_images.py doctor

1. Preview similar groups and blurry images for a folder:

python3 scripts/remove_similar_images.py analyze \
  --input-path /path/to/photos

1. Preview safe cleanup by moving similar non-keepers and blurry files into a trash folder:

python3 scripts/remove_similar_images.py analyze \
  --input-path /path/to/photos \
  --delete-similar \
  --delete-blurry \
  --trash-dir /path/to/review-trash

1. Apply the move after the preview looks correct:

python3 scripts/remove_similar_images.py analyze \
  --input-path /path/to/photos \
  --delete-similar \
  --delete-blurry \
  --trash-dir /path/to/review-trash \
  --apply

1. Permanently delete only similar non-keepers:

python3 scripts/remove_similar_images.py analyze \
  --input-path /path/to/photos \
  --delete-similar \
  --apply

Main Arguments

• --input-path: source image file or directory.
• --no-recursive: scan only the top-level directory.
• --extra-extension: include additional suffixes not covered by default.
• --limit: cap the number of scanned files for quick tests.
• --hash-method: phash, dhash, ahash, or whash.
• --hash-size: larger hashes are stricter and slower.
• --similar-threshold: maximum Hamming distance considered similar.
• --blur-threshold: Laplacian-variance cutoff for blurry images.
• --keep-policy: choose the keeper in each similar group with best, largest, newest, or oldest.
• --delete-similar: mark non-keeper files in similar groups as removal candidates.
• --delete-blurry: mark blurry files as removal candidates even when they are unique.
• --trash-dir: move files into a review directory instead of permanently deleting.
• --apply: execute removals or moves. Without this flag the script only reports.
• --report-json: save a machine-readable report for later review.
• --print-json: print the full report as JSON to stdout.

Default Heuristics

• Default similarity detection uses phash with hash_size=8 and similar-threshold=5.
• Default blur detection uses a variance-of-Laplacian cutoff of 100.0.
• Default keep-policy=best prefers non-blurry images, then sharper images, then larger images.
• Similar groups are connected components: if A is close to B, and B is close to C, they are treated as one group even if A and C are slightly farther apart.

Usage Notes

• Review the preview before adding --apply.
• Use --trash-dir for the first pass on any valuable photo collection.
• Lower --similar-threshold to be stricter. Raise it when near-duplicates are being missed.
• Lower --blur-threshold if too many acceptable images are marked blurry. Raise it when obvious blur is missed.
• Expect format support to follow Pillow and OpenCV availability in the local environment.

Output

• Text mode prints scan counts, similar groups, blurry images, and planned or applied actions.
• JSON mode includes per-image metadata, unreadable files, similar groups, planned actions, and action results.

Script

• scripts/remove_similar_images.py

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tiangong-ai
• Source: tiangong-ai/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.