Similar Textures Finder

Desktop app for finding duplicate and near-duplicate texture images in a folder. Choose a library, run a scan, and review grouped results with thumbnails, similarity scores, and per-pair match reasons.

Built with Tauri + Svelte. Supports PNG, JPEG, WebP, BMP, GIF, and TIFF.

Screenshots

Main view — pick a folder, adjust settings, and start a scan.

Results view — browse grouped matches with scores and metric breakdowns.

How it works

Scanning runs in three stages:

1. Discover — recursively collect image files under the chosen folder (symlinks skipped, paths deduplicated).
2. Ingest — for each file, compute a byte hash, decode the image, optionally preprocess it, and extract visual features.
3. Compare & cluster — link byte-identical files by hash, then compare remaining pairs with visual metrics. Connected images are grouped together for review.

Each pair can match for one of two reasons:

Reason	Meaning
Hash	Files are byte-identical (same content, possibly different names).
Composite	Files differ on disk but score above the similarity threshold on enabled visual metrics.

Groups can contain both hash and composite links (Mixed). Singleton groups (one image with no matches) can be hidden in Results.

---

Algorithm & settings

Open Settings in the app to tune detection. All values are persisted between sessions.

Stage 1 — File hash (exact duplicates)

Before any visual comparison, each file is hashed from raw bytes. Files with the same digest are grouped immediately — no decode required for the match itself.

Setting	Default	Description
Hash algorithm	sha256	Digest used for byte-identical detection. Also supports md5 and sha1.

Pairs already linked by hash are skipped during visual comparison.

Stage 2 — Preprocessing

Before features are extracted, each decoded image can be transformed:

Setting	Default	Description
Alpha crop	off	Trim transparent borders using pixels with alpha above the threshold. Useful for sprite sheets and UI icons on transparent canvases.
Alpha threshold	0.05	Pixels with alpha below this value are treated as transparent when cropping.
Rotations	off	Also extract features for 90°, 180°, and 270° rotations. Catches the same texture saved in a different orientation.
Flip	off	Also extract features for a horizontal flip (and flipped rotations when rotations are enabled).
Max decode dimension	none	Optionally downscale large images on decode to reduce memory use.

When rotations or flip are enabled, each image produces multiple feature sets. Pair scoring uses the best score across all transform combinations.

Stage 3 — Visual metrics

Three metrics can be enabled independently. Each produces a normalized score from 0 to 1 (higher = more similar). Enabled metrics are combined into a composite score using configurable weights.

pHash (perceptual hash)

Fast fingerprint of overall image structure. The image is converted to grayscale, resized to 32×32, transformed with a 2D DCT, and reduced to a 64-bit hash from the low-frequency block.

Setting	Default	Description
Enable pHash	on	Turn perceptual hashing on or off.
pHash weight	0.35	Contribution to the composite score.
pHash max distance	10	Maximum Hamming distance (0–64) for a pHash pair to count as valid. Lower = stricter.

Good for: near-duplicates, re-exports, minor compression changes. Weak against: small localized differences on an otherwise identical layout (e.g. different icon on the same template).

SSIM (structural similarity)

Compares luminance structure after both images are resized to a square. Measures how alike the overall brightness patterns are.

Setting	Default	Description
Enable SSIM	on	Turn structural similarity on or off.
SSIM weight	0.45	Contribution to the composite score.
SSIM threshold	0.90	Minimum raw SSIM for the metric to be valid. Acts as a veto: if SSIM is enabled and raw SSIM is below this value, the pair is rejected regardless of composite score.

Good for: images that look structurally alike but differ in format or minor edits. The veto helps filter pairs where structure clearly does not match.

Histogram (color distribution)

Builds a 3D RGB histogram from the resized image (transparent pixels skipped) and compares distributions.

Setting	Default	Description
Enable histogram	on	Turn histogram comparison on or off.
Histogram weight	0.20	Contribution to the composite score.
Resize size	256	Side length (px) used when resizing for SSIM and histogram extraction.
Histogram bins	512	Total bin count; converted to per-channel cube root internally.
Histogram method	correlation	Comparison method: correlation (Pearson, mapped to 0–1) or bhattacharyya.

Good for: textures with similar palettes even when layout differs slightly.

Composite score & threshold

The composite score is a weighted average of all enabled metrics that passed their validity checks:

composite = (w_phash × score_phash + w_ssim × score_ssim + w_histogram × score_histogram)
            / (sum of weights for valid metrics)

Setting	Default	Description
Threshold	0.85	Minimum composite score to link two images. Raise to reduce false positives; lower to catch more near-duplicates.

At least one visual metric must be enabled.

Results display

Setting	Default	Description
Hide single-image groups	on	Do not show groups that contain only one unmatched image.

Not yet available

ORB feature matching (enable_orb) is reserved in settings but not implemented in this build.

---

Presets

Settings can be reset quickly with Fast, Balanced, or Strict presets in the Settings dialog.

Balanced (default)

All three metrics enabled with moderate thresholds. Best starting point for most texture libraries.

	Value
Metrics	pHash + SSIM + histogram
Threshold	0.85
Weights	pHash 0.35, SSIM 0.45, histogram 0.20
Transforms	none

Use when: general-purpose deduplication — you want a reasonable trade-off between speed and accuracy without tuning individual knobs.

Fast

Skips SSIM and transform checks; uses a higher threshold and caps decode size for speed.

	Value
Metrics	pHash + histogram only
Threshold	0.92
Weights	pHash 0.55, histogram 0.45
Max decode dimension	512 px
Transforms	none

Use when: scanning very large folders where runtime matters more than catching subtle structural differences. Expect fewer matches and faster scans. Good for a first pass to find obvious duplicates.

Strict

Enables all metrics with tighter per-metric gates, plus alpha crop and rotation checks.

	Value
Metrics	pHash + SSIM + histogram
Threshold	0.78
Weights	pHash 0.35, SSIM 0.45, histogram 0.20
pHash max distance	12
SSIM threshold	0.88
Transforms	alpha crop + rotations

Use when: you need the most thorough comparison — transparent sprites, orientation variants, or libraries where Balanced misses matches. Slowest preset; run on smaller batches or overnight on large trees.

Tip: If Strict produces too many false positives (e.g. same layout, different small artwork), raise Threshold or SSIM threshold manually. If Balanced misses legitimate near-duplicates, try Strict or lower Threshold slightly.

---

Requirements

• Node.js + npm
• Rust toolchain (rustup, cargo)
• Tauri prerequisites for your OS (platform setup guide)

Run

cd app
npm install
npm run tauri dev

The Vite dev server uses port 1440. If startup fails, check that nothing else is bound to that port.

Build

cd app
npm install
npm run tauri build

Installers and bundles are written to app/src-tauri/target/release/bundle/.

Checks

cd app
npm run check
cd src-tauri && cargo test