MetricMogul/single-image-dino-patch-matcher-and-cutout

Single-image DINOv3 patch matcher

Click any patch in an image. The app finds all patches with similar DINOv3 features in the same image and exports them as a transparent PNG.

What it's for

• —Single-image co-segmentation / part discovery
• —Probing what DINOv3 considers "the same thing" without training anything
• —Building cutout datasets without a separate segmentation model
• —Sanity-checking dense self-supervised features on your own data

How it works

1. 1.The image is resized so its long side matches the target (default 448 px), then rounded down to a multiple of the patch size (16). No padding — padded regions would match as "empty area" and pollute the similarity map.
2. 2.The image is passed through facebook/dinov3-vits16-pretrain-lvd1689m. CLS and the 4 register tokens are stripped; only patch tokens remain.
3. 3.Patch tokens are L2-normalized so cosine similarity is just a dot product.
4. 4.When you click a patch, its token is used as a query against all other patches. Patches that are both in top-k and above the threshold are kept.
5. 5.Kept patches are highlighted as a red mask. Cut selected patches exports them as a transparent PNG, cropped to the mask's bounding box.

The mask is the raw set of matched patches — no CRF, no connected components, no morphology. What you see is what DINOv3 actually represents.

Notes and limitations

• —Edges are stepped because the mask lives on the 16×16 patch grid. This is a feature, not a bug — it shows you DINO's actual resolution.
• —Top-k and threshold are combined (intersection), so tightening either shrinks the selection. Top-k controls how aggressive the match is; threshold controls how strict the similarity gate is.
• —A single patch query can match semantically related but visually different regions (e.g. clicking on one wheel often selects all wheels). This is the whole point of self-supervised dense features.

Model

facebook/dinov3-vits16-pretrain-lvd1689m — patch size 16, 4 register tokens, ViT-S backbone. Runs comfortably on CPU; faster on GPU.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference