Similarity

Similarity

RawCull can rank every image in a catalog by visual similarity to a chosen anchor image, so near-duplicate frames and series shot in similar conditions float to the top of the grid.

How It Works

Similarity ranking is powered by Apple’s Vision framework:

• A 512 px thumbnail is extracted from each ARW file — no full RAW decode is needed.
• VNGenerateImageFeaturePrintRequest produces a compact numerical embedding that describes the overall visual content of the image (colour, texture, composition).
• When you select an anchor image, RawCull computes the distance between the anchor’s embedding and every other indexed image. A shorter distance means greater visual similarity.
• If sharpness scoring has already run, subject labels from saliency analysis are used to apply a small penalty when two images depict different subject types. This keeps the visual embedding as the dominant signal while nudging images of the same subject type slightly higher.
• The grid re-sorts in ascending distance order — most similar images appear first, directly after the anchor.

Using Similarity

1. Open a catalog and wait for thumbnails to finish generating.
2. Click Index for Similarity in the toolbar to build embeddings for all files. A progress indicator shows how many files have been processed and an estimated time remaining. Already-indexed files are skipped if you re-index.
3. Select the image you want to use as the reference and click Find Similar (or the equivalent toolbar button). RawCull sets that image as the anchor and calculates distances to all other indexed images.
4. The grid sorts by similarity automatically — the most visually similar frames appear immediately after the anchor.
5. To return to normal sort order, toggle the similarity sort off in the toolbar or sidebar.

Indexing similarity completed.

Sort images after similarity selected image.

Tips

• Run sharpness scoring first — if subject labels are available from the sharpness analysis pass, the subject-mismatch penalty improves the ranking of same-subject bursts.
• Re-index after adding files — any file added to the catalog after the last index run will have no embedding and will not appear in similarity results. Run Index for Similarity again to include new files; already-indexed files are skipped automatically.
• Use similarity alongside sharpness — a common workflow is to score sharpness first to find the sharpest frame in a burst, then use that frame as the anchor to group all related near-duplicates for quick review and culling.
• The anchor image always scores zero distance — it is the reference point, so it appears at the top of the sorted list regardless of its sharpness score.