Annotension

Annotate similar
images together.

Clustering groups images by visual similarity, so you can work through related scenes together and inspect unusual images sooner.

Use it before labeling your starting dataset or when reviewing a new batch of model predictions. It organizes the work; it does not create boxes or approve labels.

Spend less time switching context

Moving between similar images can make it easier to apply the same box boundaries and class rules. When reviewing predictions, look for a recurring mistake across the group, such as missed small objects or loose boxes.

Inspect examples from different groups when deciding what to annotate next. This can help you notice scenes that are underrepresented in your starting labels. Clusters do not measure class balance or guarantee useful training coverage.

Similar-looking images arranged in groupsAn illustration of two groups of image thumbnails. Each group contains related views. Visual similarity is not a class label.Group AGroup B
Illustration: visual groups help organize inspection. A cluster is not an object class.

Cluster your gallery

  1. Open the gallery

    Choose Cluster images….

  2. Choose the grouping

    Select an encoder, group count, and effort.

  3. Work group by group

    Enable Group by cluster and filter Needs review.

  4. Inspect unusual images

    Enable Outliers first, then review each image.

Open an image to annotate it, then return with ⌘ G. Click a cluster header, or use the compact Cluster navigation selector, to jump between groups.

Choosing settings

Apple Vision is the default. Vision 2 needs macOS 14 or later; Vision 1 is the older alternative. MobileNetV4 Conv-Medium is available when included in your installed build. These encoders run on your Mac without Python, image uploads, or user-time model downloads. MobileNetV4 uses a center crop that can omit edge objects.

Choose 2–256 groups, bounded by the image count, and Fast, Balanced, or Thorough effort. Start with a manageable number of groups and inspect whether they help. More effort does not guarantee better visual groups; identical image features can produce fewer groups than requested.

Features are cached, so changing the group count or effort reuses the extracted features. Completed regrouping preserves annotations and review states and supports Undo/Redo. Cancellation or failure keeps the preceding groups.

Use outliers as inspection cues

A high outlier score means an image is visually far from the center of its group. It may be worth checking for a different viewpoint, background, or scene before continuing with the rest.

The score is not model confidence, an anomaly probability, or evidence that a label is wrong. Inspect the image before deciding what it needs.

Keep review explicit

Grouping does not copy labels between images, accept model suggestions, or mark images reviewed. Check every image, resolve pending suggestions, and approve it yourself.

Clustering also does not create train/validation splits or prevent leakage between related frames. Keep inspecting split coverage and near duplicates separately.

Continue the local loop →