🔎 FeatLens — see what any vision model encodes

Render feature maps from any layer of any vision model, colored by PCA, cosine-similarity to a seed patch, k-means segmentation, or a foreground mask. In cosine mode, click the image to move the seed patch.

GitHub · PyPI · Docs

The first render of a model downloads its weights (a few seconds for the default ViT-S); repeat renders on the same image are cached.