AI Image Similarity Detector — free online photo comparison & duplicate finder
Upload two images and a five-layer pipeline compares them with MobileNet deep features, perceptual hash (pHash), color histogram, SSIM and pixel MSE — accelerated by WebGL and WebAssembly. Detects identical, near-duplicate, cropped, resized, recolored and semantically-similar images. No upload, no signup, everything runs in your browser.
Running 5-layer image similarity analysis…
—
Compare two images to see the verdict.
When to run an image similarity check
Compare two images whenever you need to know whether they're the same, near-duplicates, or just somewhat alike — works for photos, screenshots, artwork, product shots and AI-generated images.
Find duplicate photos
Spot duplicate or near-duplicate shots in a photo library, asset folder or product catalog.
Reverse image checks
Compare a found image to your original to confirm a match before filing a copyright notice.
Detect crops & resizes
Match a thumbnail or cropped version against the full original — pHash and deep features handle both.
Compare recolours
Color histogram and SSIM tell you if a recoloured or filtered version still matches the original.
AI image audits
Compare an AI-generated image to a reference photo to gauge how closely the model copied it.
Evidence for claims
Document deep-feature, pHash and SSIM evidence for DMCA, marketplace or moderation workflows.
How to compare two images (4 steps)
Upload Image A
Drag-drop, paste (Ctrl/Cmd+V), or click the left card to pick the first image. URL load is also supported.
Upload Image B
Drop the second image into the right card.
Compare
Click Compare images. The detector runs six stages — load, decode, pHash, deep features, histogram + SSIM, final assessment — in a couple of seconds.
Review & export
Read the verdict tier, six metrics, side-by-side preview and detection signals. Export a TXT or HTML/PDF report when you need evidence.
Frequently asked questions
An image similarity detector compares two images and outputs a similarity score. HCODX's detector uses a 5-layer pipeline: MobileNetV2 deep features for semantic similarity, perceptual hash (pHash via DCT) for near-duplicate detection, color histogram (Bhattacharyya distance) for color distribution match, SSIM for structural similarity, and pixel MSE for identical-image detection.
Yes. 100% free, ad-free, no signup. There are no daily limits or paywalls.
No. Images stay in your browser. Everything runs locally with TensorFlow.js (WebGL accelerated, WASM fallback). The MobileNet model is downloaded once and cached. URL inputs go directly from your browser to the hosting site.
JPG, PNG, WebP, GIF, BMP, AVIF (browser-dependent) and SVG. Animated GIFs use the first frame.
Identical images are caught at the pixel-MSE stage and short-circuit to a 100% IDENTICAL verdict — catches re-uploads of the same file regardless of filename or container.
Yes. pHash and MobileNet features are robust to resize, mild crop, compression and color adjustments — a thumbnail and a full-size version of the same photo will score very high.
IDENTICAL (≥97%) — essentially the same image. NEAR DUPLICATE (≥90%) — small edits, re-saves, watermarks or light cropping. VERY SIMILAR (≥78%) and SIMILAR (≥62%) — progressively softer matches. SOMEWHAT SIMILAR (≥40%) — shared colors or composition. DIFFERENT (<40%) — no significant similarity.
Yes. After each run you can export a plain-text report or a printable HTML/PDF report listing the verdict, all metrics, image properties and an executive summary.
About this image similarity detector
HCODX's AI Image Similarity Detector is a free, ad-free, no-signup tool that compares two images directly in your browser. It runs on top of TensorFlow.js with the MobileNetV2 image-classifier as a feature extractor, plus custom JavaScript implementations of pHash, color histograms, SSIM and pixel MSE running in a Web Worker. Nothing is uploaded.
The 5-layer pipeline
- Layer 1 — Deep features (MobileNetV2). A 1024-dim embedding is extracted from each image. Cosine similarity captures semantic similarity — what the picture is of — regardless of resolution, crop or small edits.
- Layer 2 — Perceptual hash (pHash). 32×32 grayscale → 2D DCT → top-left 8×8 → 64-bit hash. Hamming distance gives near-duplicate detection that survives resizing, mild cropping and compression artefacts.
- Layer 3 — Color histogram. RGB binned into 8×8×8 = 512 bins, L1-normalised. Bhattacharyya coefficient measures how alike the color distributions are.
- Layer 4 — SSIM (Structural Similarity Index). Sliding-window comparison of luminance/contrast/structure between the two images — robust to lossy compression.
- Layer 5 — Pixel MSE. The hard pixel-difference baseline. Triggers the IDENTICAL shortcut when near zero.
Final score
The five signals are blended with weights that favour deep features (the most accurate semantic signal): 0.35 · deep + 0.25 · pHash + 0.20 · SSIM + 0.10 · histogram + 0.10 · (1 − MSE). Identical inputs (MSE ≈ 0 + pHash distance 0) shortcut to 100%.
Verdict bands
≥ 97%— IDENTICAL. Same image bytes or pixel-perfect re-encode.≥ 90%— NEAR DUPLICATE. Light edits / watermark / mild crop / re-save.≥ 78%— VERY SIMILAR. Same subject, different framing.≥ 62%— SIMILAR. Shared content or strong style overlap.≥ 40%— SOMEWHAT SIMILAR. Comparable composition, colors or category.< 40%— DIFFERENT. No meaningful similarity.
Performance
Backend preference: WebGL → WASM → CPU. MobileNet inference on a 224×224 input runs in tens of milliseconds on WebGL. The heavy O(W·H) loops for SSIM, histogram and pHash are dispatched to a Web Worker so the main thread stays responsive.