NCD · Normalized Compression Distance

Two things are the same structure when they compress better together than apart.
A model-free similarity measure. It extracts no features and is told nothing about what to look for — it asks only whether a compressor finds shared structure across two objects. Identical inputs read ~0; objects with nothing in common read ~1. The number means nothing in isolation, so this tool computes three reference anchors alongside yours and shows you where your pair lands among them.
Image A drop, paste, or click to choose
Image B drop, paste, or click to choose
normalize to
NCD ( A , B )
where your pair sits
Computed live from your two images. Three anchors fix the axis: A vs itself (zero), A vs shuffled-A (same pixels, structure destroyed), and A vs random noise (far end). Your A↔B distance is placed among them — scaled to these images, not an abstract 0–1. If A↔B reads near or past the shuffled anchor, the images differ in spatial structure, not just content.
C(A)
compressed bytes
C(B)
compressed bytes
C(A·B)
compressed together
NCD = [ C(A·B) − min( C(A), C(B) ) ] / max( C(A), C(B) )
both concatenation orders averaged · large-window LZ · no compressor window ceiling