PreviewCandela is still in development, with capabilities current as of August 23, 2026.

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.

Two benchmarks

One public, one built for the work we protect.

No version of Candela trains on either benchmark's corpus.

AnyPattern

Public

25,000 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.

CopyNCE (ViT-S)open research, newest copy-detection model
0.269
candela-2-pilotthe single model serving Lantern today
0.229
candela-1.3the previous ensemble, for comparison
0.093
candela-1.2the earlier ensemble, for comparison
0.058
SSCDopen copy-detection specialist
0.056
DINOv3 (ViT-L)open research, self-supervised
0.005
OpenCLIP (ViT-L)open research, vision-language
0.002

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

Ours

492 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.

candela-2-pilotthe single model serving Lantern today
83%
candela-1.3the previous ensemble, for comparison
77%
CopyNCE (ViT-S)open research, newest copy-detection model
74%
candela-1.2the earlier ensemble, for comparison
74%
DINOv3 (ViT-L)open research, self-supervised
64%
SSCDopen copy-detection specialist
64%
OpenCLIP (ViT-L)open research, vision-language
54%

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.

Small by design

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.

candela-1.3 (retired ensemble)models fused; 2,048 numbers per fingerprint across three weighted vectors
~350M parameters
DINOv3 (ViT-L)open research, self-supervised; 1,024 numbers per fingerprint
~300M parameters
CopyNCE (ViT-S)leading open model in our tests; 512 numbers per fingerprint
~22M parameters
candela-2-pilot (serving today)one 512-number fingerprint per artwork
~22M parameters