How to group your pictures
- Add your pictures. Drop a folder or many pictures anywhere on the page, or click the box, or use Choose a whole folder instead. Subfolders are included, and files that are not pictures are skipped. JPEG, PNG, WebP and iPhone HEIC photos work.
- Wait while they are read. The first time, the page downloads the model (see below). The groups appear when the last picture is read.
- Set how alike pictures must be. Move the slider, or use a quick setting. The groups change at once, and the line above them says how many groups there are and how many pictures are on their own.
- Look at the groups. Each shows its pictures, with the most typical one first. Click a picture to open the original. The picture with the most pixels in each group is marked Largest.
- Save the result. Choose one of the three downloads (below).
Nothing is moved or deleted: the page cannot touch your folder. It makes ZIPs and a list for you to use.
The slider and the quick settings
The slider runs from Same kind of thing to Nearly the same picture. Pictures are joined into a group while they are at least that alike. The three quick settings are good starting points. Here is what each did on 80 photographs of ten kinds of thing (eight each of cats, dogs, beaches, mountains, cars, flowers, pizza, bridges, birds and trains):
| Setting | What it joins | On our 80 photos |
|---|---|---|
| Same subject | Pictures of the same kind of thing: dogs with dogs, trains with trains | 13 groups holding 73 of the 80 pictures; 84 of every 100 pictures in a group were of the kind that was most common in that group |
| Same scene | Pictures of the same place, object or event: a burst of shots, or the same car from two sides | 13 groups holding 41 pictures; 98 of every 100 matched their group's main kind; 39 pictures were left on their own |
| Near-copies | The same picture saved again: resized, mirrored, slightly cropped or re-compressed | 3 pairs of very alike photos among the 80 (all of the same kind); the other 74 were left on their own |
The slider starts a little left of Same scene. A picture that is not like any other stays on its own: those are listed last, in one card.
Finding near-copies, and keeping one
To clean up copies, choose Near-copies. We tested it on 40 different photos plus 10 copies of them (four resized to half, four mirrored, two re-saved as a JPG at quality 35): all 10 copies were grouped with their originals, and no group joined two different photos. A copy that has been changed a lot, such as a strong crop, big color changes or a very low JPG quality, can be less alike than that, so if you see copies sitting alone, move the slider a little to the left.
Then use Download one picture from each group. It gives you the picture marked Largest from every group, plus every picture that was alone, in one flat ZIP. That is the list of what to keep. The Download the list (CSV) button gives the same answer as a table, with a keep column of yes and no, which you can sort in a spreadsheet to find the pictures you can delete yourself.
The duplicate finder does a similar job with a different method, comparing the shapes in the picture, which suits exact copies and needs no model download. This page works by what the pictures show, so it also puts together pictures of the same scene that are not copies. That is useful, and it also means you should look at a group before you delete anything from it. Our guide to finding duplicate photos explains the difference.
What you can save
- A ZIP with a folder for each group (
grouped-pictures.zip): folders namedgroup-01,group-02and so on, with the biggest group first, and one folderon-their-own. The files are your originals, unchanged. - One picture from each group (
one-per-group.zip): the Largest of each group, and the pictures that were alone. - The list (
groups.csv): columnsfile,groupandkeep, one row per picture, with the folder path when you added a folder.
How good is it?
It groups by what a picture looks like to a model that sees a 224 × 224 square from the middle of each picture, so something at the edge of a wide photo is not seen, and a very different crop of the same scene may land elsewhere. It mixes up things that look alike: in our 80 photos the common mix-ups were dogs with cats, bridges with mountains, and cars with trains. It is not made to tell people apart and should not be used for that. Use it to sort folders and spot copies, and look at the groups yourself.
Speed, the model and the download
The page uses CLIP (the ViT-B/32 model) from OpenAI, in a compact 8-bit form of the files converted by the transformers.js project. It turns each picture into 512 numbers, and pictures that look alike get numbers that are close. This page needs only the picture half of the model: about 85 MB the first time, kept by your browser. It runs in this tab through transformers.js and ONNX Runtime. OpenAI's CLIP repository is under the MIT license.
Reading takes about a quarter of a second per picture on a laptop (80 pictures in about 20 seconds on an Apple M1 Max laptop that was busy with other work). The page groups up to 4,000 pictures at a time, because it compares every picture with every other. OpenAI's model card for CLIP warns that it struggles with fine-grained classification and counting, and found significant disparities with respect to race and gender, so do not use the groups to judge people.
Privacy
The grouping happens in this tab. For what the site collects, see our privacy policy.
Frequently asked questions
What is the difference between this and the duplicate finder?
The duplicate finder compares the shapes in each picture, which finds exact and resized copies quickly. This page compares what the pictures show, so it also groups pictures of the same scene or the same kind of thing, but it needs a model download and it needs more care before you delete anything.
How many groups will I get?
It depends on your pictures and the slider. Moving the slider to the left makes fewer, bigger groups, and to the right more, smaller ones. The line above the groups always says how many there are.
Will it delete my duplicates?
No. A web page cannot change your folder. It gives you a ZIP with one picture from each group and a list saying which to keep, and you decide what to do with the rest.
Why are two pictures of the same place in different groups?
The model sees only the middle square of each picture, and two shots of one place can look quite different there. Move the slider to the left to join more loosely.
Can it group photos by person?
It is not made for that and we do not recommend it: CLIP is not a face recognition model, and its authors warn against using it to identify people.