Asset Similarity

Open the workflow

Select an image in a document, open the Intelligence panel, and choose Similar or Near duplicates. The first use may offer an explicit download for the optional local model.

Which mode should I use?

• Similar searches compatible semantic image vectors and does not hash-prefilter candidates.
• Near duplicates favors exact content identity and visual fingerprints. It is the better choice for exports, resized copies, and lightly edited variants.

Privacy and limitations

Inference runs locally in a worker once the model is installed. The current workflow scans up to 30 images in the open document. It does not support text prompts, persistent cross-project indexing, automatic clustering, or automatic deletion.

Model notes

Varve's image-to-image lane uses the checksum-pinned DINOv2-small ONNX encoder, selected by measured retrieval quality, speed, and download size on a Varve-specific corpus (reference-vector parity verified against an independent runtime). The natural-language lane pairs the SigLIP text graph with SigLIP image embeddings. Model changes require reference-vector parity, retrieval-quality evidence on the Varve corpus, and a license/distribution review.