---

name: imagededup description: "Route image deduplication tasks for the imagededup package: hash-based duplicate detection, CNN-based duplicate detection with custom models, and evaluation or plotting of duplicate maps." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

imagededup

Use this skill for the imagededup package when the task is about finding exact or near-duplicate images, generating encodings, comparing duplicate maps, or plotting duplicate relationships.

Start here

• Use sub-skills/hashing/SKILL.md for perceptual, average, difference, or wavelet hashing; duplicate search by Hamming distance; or remove-list generation.
• Use sub-skills/cnn/SKILL.md for CNN-based encodings, pretrained or custom models, cosine-similarity duplicate search, or GPU/CPU behavior.
• Use sub-skills/evaluation/SKILL.md for evaluating retrieved duplicate maps and plotting duplicate groups.

What this skill covers

• Encoding one image or a directory of images.
• Finding duplicates from a directory or from a precomputed encoding map.
• Removing duplicate filenames with a heuristic list, not deleting files.
• Evaluating retrieved duplicate maps against a symmetric ground-truth map.
• Plotting one image and its duplicates.
• Custom CNN model wrappers and the bundled pretrained CNN backbones.

What this skill does not cover

• Package release automation.
• Docs building or site generation.
• Benchmark reproduction as a maintainer task.
• External dataset downloads such as the CIFAR10 Colab example.

Quick install for inspection or verification

Create an isolated Python 3.11 environment, then install the package from PyPI plus the runtime imports that this version uses but does not fully declare in metadata.

python -m pip install "imagededup==0.3.3.post2" numpy scipy

If you plan to run package tests in a separate source checkout, also install pytest and pytest-mock there.

Minimal import check

from imagededup.methods import PHash, AHash, DHash, WHash, CNN
from imagededup.evaluation import evaluate
from imagededup.utils import plot_duplicates, CustomModel

Routing guidance

Read the matching sub-skill when the user asks for any of the following:

• exact or near duplicate lookup with PHash, AHash, DHash, or WHash
• hash generation from files, directories, or image arrays
• find_duplicates / find_duplicates_to_remove with Hamming thresholds
• search_method, bktree, brute_force, or brute_force_cython
• CNN encodings, cosine similarity, or CustomModel
• pretrained MobileNetV3, ViT, or EfficientNet backbones
• evaluate, MAP, NDCG, Jaccard, precision, recall, or F1
• plot_duplicates

Common first checks

1. Confirm the image paths or encoding map are valid.
2. Confirm the threshold type matches the workflow.
3. Decide whether the task is hash-based or CNN-based.
4. Use evaluation only after you already have a ground-truth map and a retrieved map.
5. Use plotting only after the duplicate map is non-empty.

Helpful bundled references

• Read references/installation.md for repo-specific install notes, including missing metadata dependencies and CNN weight downloads.
• Read references/troubleshooting.md for cross-cutting import, image-format, backend, and Windows multiprocessing issues.
• Read references/repo-provenance.md when you need the source revision behind this skill.
• Read references/repo-routing-metadata.json when another agent needs to route tasks into this repo skill.

Script helpers

• Run sub-skills/hashing/scripts/hash_smoke.py to generate a synthetic hash workflow smoke check.
• Run sub-skills/cnn/scripts/cnn_smoke.py to exercise CNN encoding and duplicate search with a safe synthetic fixture.
• Run sub-skills/evaluation/scripts/evaluate_plot_smoke.py to exercise metrics and duplicate plotting with generated data.

Fast routing summary

• Hashes and Hamming distance -> hashing sub-skill.
• CNN encodings and custom models -> CNN sub-skill.
• Metrics and plots -> evaluation sub-skill.

When in doubt, prefer the most specific sub-skill that owns the user-visible workflow, then use the shared references only for install or troubleshooting context.