Numbers matter.
Candela is measured on both AnyPattern, a public, photography-centered benchmark, and LUX26, a Lantern-developed benchmark that reflects the platform's digital art focus and specialization.
One public, one built for the work we protect.
No version of Candela trains on either benchmark's corpus.
AnyPattern
Public25,000 altered copies / 1,000,000-image gallery
Each copy is altered by one of ten transformation patterns held out of training, then searched against a million unrelated images. Scored by µAP, which asks whether the true original outranks every decoy across the whole result set.
AnyPattern is an open, photography-centered copy detection benchmark. candela-1.3 maintains weights that favor digital art and is greatly outperformed by CopyNCE on its own. These figures are not comparable to the original, published scores as those baselines were retrained on its own transformation patterns while our runs had not.
candela-2-pilot, the art-specialized model, is serving today.
LUX26
Ours492 registered artworks / 5,400 unrelated works / 8,856 altered copies
Openly licensed illustration, comics, concept art, and vector work: 492 pieces registered, then altered 18 ways (crops, rotations, recolors, screenshots, memes, print scans, AI re-edits). The decoys are deliberately hard, drawn from the same artists and art styles. Recall is measured at a fixed 1% FP rate.
LUX26 is a digital art-centered benchmark that requires models to perform both a detection and verification test to determine a recall percentage. Corpus and methodology will be released soon.
candela-2-pilot, the art-specialized model, is serving today.
A fraction of the size, built for one job.
candela-2 is a single compact model specialized for art, next to general-purpose open models many times its size. Smaller means cheaper to run, faster to answer, and practical to serve on modest hardware.