Why we built it.
General image models are trained on photographs and built to answer broad questions. Candela is only asked one. The version serving traffic today reaches that focus by learning on art directly; the ensemble it replaced reached it by fusing three open models and tuning them against art.
The one question
Are these two images the same work?
Same work, altered
A crop of a screenshot of a filtered repost. An AI-assisted edit of a registered piece. The file itself may have changed, but the work does not. Candela provides a confidence score it has earned on held-out evaluations.
Same style, different work
Another artist working in a similar style, palette, or genre may be confused for copying by general models. Candela is built not to do this, so style alone never produces a match.
The output
One number.
Candela returns a similarity score. The thresholds below are the ones running in production, tuned against measured outcomes, and they decide how far a result is allowed to travel. A match found only by the semantic signal is dropped rather than surfaced.
Geometric verification, which checks that individual points agree on one alignment, is in development. That overlay is what a takedown notice will carry.
All digital media, not just photography.
Line work, flat color, cel shading, deliberate composition: art behaves differently from photography, and models tuned on photos underperform on it. The art-specialized model now serving learned on public-domain and licensed art with the alterations the internet actually applies.
One answer, many doors.
Candela powers Lantern first. But whether two images are the same work is a question platforms, marketplaces, and model builders all need answered, and Candela is built to answer it directly, as a service, at scale.