Image Similarity
Compare images and find near-duplicates using CLIP embeddings — runs on your device.
About Image Similarity
Image Similarity compares two or more images and tells you how visually alike they are, using CLIP embeddings computed right in your browser. It is built for spotting near-duplicates — resized copies, re-saved versions, lightly edited variants — where a byte-for-byte file comparison would miss the match. Nothing is uploaded; the model runs on your device and returns a similarity score as JSON.
- Category
- inspect
- Input
- Accepts: image/jpeg, image/png or image/webp.
- Output
- Outputs: application/json.
- Cost
- Free, runs in your browser
- Memory
- medium
- Install group
- image-ai
Common uses
- Find near-duplicate photos in a folder before you cull a bloated camera roll
- Check whether a thumbnail and a full-size asset are actually the same image
- Detect a lightly cropped or re-exported copy of a stock image you already own
- Group visually similar product shots when deduplicating a catalog
- Verify that a compressed or resized version still matches the original closely enough to swap in
- Spot reused imagery across a set of submissions or uploads
Frequently asked questions
What does the score mean?
It returns a similarity value as JSON based on CLIP embeddings. Higher means more visually alike; near-identical images score very high, unrelated ones low.
Is this the same as a pixel diff?
No. A pixel diff needs identical dimensions and flags every changed pixel. This compares the images semantically, so a resized or re-encoded copy still registers as highly similar.
Do the images leave my device?
No. The CLIP model runs locally in your browser, so your images and the resulting scores stay on your machine.
Which formats are supported?
JPEG, PNG, and WebP.
Do the images have to be the same size?
No. Unlike a strict pixel comparison, embedding-based similarity works across different dimensions and aspect ratios.
Will it match two photos of the same object taken differently?
It can register them as related because it compares visual content, but lighting and angle changes lower the score. It is strongest at catching near-duplicates of the same underlying image.
Keywords
- similarity
- duplicate
- compare
- clip
- perceptual
- near-duplicate
- embedding