--- title: "Image Similarity Search - Find Local Photo Matches" description: "Find similar images in your own photo collection with local AI. Choose a reference, compare up to 30 pictures, and export ranked results without uploading files." url: https://candoya.com/tools/image-to-image-search/ locale: en type: browser-tool --- # Image Similarity Search Category: [Social & Creator](https://candoya.com/tools/social/) ## What is image similarity search? Image similarity search ranks a collection of pictures by their resemblance to a reference image. This tool compares AI image embeddings on your device, then shows the closest matches first. Choose your own collection of up to 30 images; the search stays within those files and does not search the web. The CLIP ViT-B/32 image encoder creates a 512-number embedding for each image. Cosine similarity compares their directions and ranks the collection from highest to lowest. The score is not a probability, a copyright check or proof of an exact duplicate. Two different photos of the same subject can score highly. Previews are resized to at most 1,024 pixels on the longest side. For comparison, each original image is center-cropped to a square, composited onto white and resized to 224 by 224 pixels. Images, previews and results remain in this tab; Clear all releases the selection. Model and runtime downloads can remain in the browser cache. ## How to use it 1. **Choose a reference** - Choose a JPG, PNG or WebP reference image, or load the example collection. 2. **Add a collection** - Add the local images you want to search. Remove any image you do not need. 3. **Run the search** - Select Find similar images. The first run downloads the model; subsequent runs may reuse the browser cache. 4. **Review and export** - Compare the ranked thumbnails and cosine scores, then copy or download the filenames and scores as CSV. ## Who it's for - Find another product photo with a similar subject in a small asset collection. - Sort a shortlist of travel photos by resemblance to a reference scene. - Compare illustration or mood-board candidates without sending the files to a search service. ## Expert note Center cropping changes what the model sees: a subject near the edge of a wide panorama may disappear from its input. Crop around the subject before comparing when edge details matter. Inspect the thumbnails yourself, especially when every candidate is unrelated or when tiny text and fine details distinguish the pictures. ## Frequently asked questions ### Are my images uploaded? No. Image decoding and comparison run in your browser. The first search requests the AI model from Hugging Face and the runtime from jsDelivr, which receive ordinary download requests. Your selected image files and their embeddings are not sent to those services or to CanDoYa. ### Is image-to-image search free? Yes. This local collection search is free and does not require an account. Your browser performs the computation, so speed depends on your device. The first search downloads about 64 MB of model weights plus the runtime; a cached model can make later searches quicker. ### How many images can I search? Choose one reference and up to 30 collection images. Each file must be JPG, PNG or WebP, at most 10 MB, 12 megapixels and 8,192 pixels per side. The combined file size limit is 100 MB. Smaller collections work better on phones and older devices. ### Does this find images on Google or across the internet? No. Results come only from the collection you select on your device. This is useful for finding similar pictures among files you already have. It cannot locate a picture's original website, identify its owner or discover copies elsewhere on the internet. ### What does the cosine similarity score mean? The score ranges from -1 to 1; higher means the image embeddings point in more similar directions. It helps order one collection, but there is no universal score that proves a match. Similar subjects can score highly despite different colors or layouts. All selected candidates are ranked, including unrelated ones. ### Can it find exact duplicates or identify a person? It does neither reliably. This model compares broad visual and semantic features, so it is not a byte-for-byte duplicate detector, facial identity system or text reader. Review the pictures yourself. For exact file identity, compare file hashes; for small visual differences, use a direct image comparison. ### What if my browser has no WebGPU? The tool uses WebAssembly on your CPU when WebGPU is absent. If graphics initialization or inference fails, it retries the entire search on the CPU so all scores use the same backend. The model runs in a worker, and Cancel stops that worker while keeping your selected images. ## Related tools - [Image Captioner](https://candoya.com/tools/image-captioner/) - [Object Detection](https://candoya.com/tools/object-detection/) - [Depth Estimation](https://candoya.com/tools/depth-estimation/) - [Background Remover](https://candoya.com/tools/background-remover/) --- *Markdown edition of , published for AI agents and other automated readers. Index of key pages: .*