Questions, answered plainly.
Is Candela trained on my art?
Not unless you say so. Registered work is excluded by default. You can opt an individual piece in from its settings if you want to help improve detection, and you can turn that off again at any time, which removes the work from future training runs. Work already used in a completed training run cannot be removed from that model, so we only ever train on work someone actively opted in.
It is worth being clear about what Candela is, because "training an AI on your art" usually means something very different. Candela only ever answers one question: are these two images the same work? It recognises art, it cannot generate or imitate it, and it has no ability to produce an image at all. Every piece opted in makes the detector better at finding stolen work, including yours.
How is Candela trained?
The version serving traffic today, candela-2-pilot, is an art-specialized model fine-tuned from an open copy-detection base, on ordinary local hardware rather than in a data center. Its training set is public-domain museum artworks, openly licensed art, and a public research dataset of everyday photographs, which is what teaches it to survive crops, filters and re-compression. No registered work is in it unless its owner opted that piece in.
What does the confidence percentage mean?
It is a similarity score from our matching model, with the band thresholds tuned against measured outcomes on held-out evaluations. A higher score means a stronger signal that the images are the same work, and where the boundaries between bands sit was chosen from graded data. It is not a calibrated probability.
Can it tell 'inspired by' apart from 'copied'?
That is the core design goal. Candela is built to separate the same work in altered form from different works in a similar style. Style similarity alone does not produce a match, and we never treat style as infringement.
What is geometric verification?
A second stage, currently in development and not yet part of the served pipeline, finds individual points that correspond between two images and checks that they agree on one consistent transformation. It is what produces the evidence overlays shown on the overview page.
Does it detect AI-generated images?
That is a different question and a different product. Candela answers whether an image is a copy or derivative of a specific known work, including many AI-made derivatives of it. It does not judge whether an image was made by AI in general.
Can I use Candela in my own product?
Direct API access is opening gradually; see Pricing and access to request it. Platforms integrating with Lantern already get Candela through the platform kit at no cost.
Why the name?
The candela is the unit of luminous intensity: how brightly a light shines in one direction. A fitting name for the model that measures how strongly a match shines.