FAQS
We have compiled a list of the most commonly asked questions
If you can’t find the answer you’re looking for, please don’t hesitate to contact us at office@art-recognition.com – we’re here to help!
As of November 2024, Certificates of Authenticity issued by Art Recognition are accepted at auctions in Switzerland. The first auction house to sell an artwork based only on a Certificate of Authenticity from Art Recognition was Germann Auctions from Zurich.
At Art Recognition, keeping your information safe and private is a top priority. Everything from the ownership details of the item we examine to the results we provide is kept strictly confidential. We only share case studies or images with third parties or publish them if we have your written consent.
No. The Art Recognition AI System can handle any artist, provided there are enough digital images of authentic artworks by that artist to train on.
Our AI generates accurate results, even when artists undergo substantial stylistic shifts throughout their careers – a case in point being Cézanne. To effectively capture these distinct styles, we construct models on images of the artist’s specific artistic period.
Indeed, the Art Recognition AI System can deal with this situation. The unique brushstroke of an artist spans across all styles, also on periods that contain less works. Therefore, if your artwork is an authentic piece by the assumed artist, the brushstroke will be recognized.
Art Recognition possesses an extensive proprietary in-house database. Additionally, we have established collaborations with leading image providers like Scalarchives and Bridgeman Images, granting us access to a vast repository of millions of high-resolution artwork images.
There is no universal standard for the size of a training dataset, so we cannot provide a fixed number. Our aim is to build comprehensive datasets by including all validated works of a given artist. As a result, dataset sizes vary widely—some artists have fewer than 50 authenticated works, while others have thousands. The number of negative examples (non-authentic images) typically exceeds the number of authentic ones slightly. These are carefully chosen to closely resemble the authentic works in style and composition.
To work with very small datasets—such as those with only 20 to 30 images—we first augment the model by generating embeddings, which are numerical representations of images. By combining the embeddings of two authentic artworks, we create a new embedding that captures features from both and is treated as authentic by the AI model. These generated embeddings are not visual images but data points that help the model learn more effectively from limited input.
The datasets used to train the AI on different artists are cross-referenced with the Catalogue Raisonné of each respective artist. This ensures that the training dataset for any given artist includes all images officially documented as their work. For example, our Van Gogh paintings dataset contains nearly one thousand images, all verified against the La Faille Catalogue Raisonné. Similarly, the Raphael dataset, which includes approximately 100 images, has been cross-referenced with the Catalogues Raisonnés by Luitpold Dussler and Jürg Meyer zur Capellen. You can download the Raphael dataset for free at this link: https://github.com/Art-Recognition/Raphael_Dataset
Yes, whenever available, we incorporate infrared, X-ray, and ultraviolet images into our AI training. These imaging techniques reveal features such as underdrawings, pentimenti, and later interventions, which can act as unique identifiers. Our training includes such images from authenticated works by the artist, as well as from known imitations or works by other artists in a similar style. Just as AI can learn to recognize an artist’s stylistic traits, it can also analyze patterns found in these technical images, such as preparatory sketches and revision marks.
Our approach is based on AI technology that learns the patterns and characteristics of an artist’s work and compares these with the artwork in question to authenticate it. Therefore, we do not carry out any materia analysis on the artwork.
Provenance research is not incorporated into the AI analysis. The analysis conducted by AI does not include tracing the origin or history of the items being analyzed.
The fees vary according to the technical complexities of the artwork and the artist. Certain artists or paintings might require more extensive efforts in collecting and curating training images, as well as in analyzing the outcomes. Furthermore, the costs can be impacted by your individual requirements; for example, if you request an analysis using a model trained solely on a specific phase of the artist’s creative output, the fees could potentially increase.
Definitely! Write us at office@art-recognition.com in order to receive a package tailored to your specific request.
Normally we deliver the Art Recognition Certificate or Report within 7 to 10 working days. If we need longer, we inform you right away.
Certainly. We require a short list of two or three names of presumed artists, as we will be conducting the analysis for each artist. We are happy to offer you a tailored commercial proposal for your package analysis.
Absolutely. Please make sure that the photo is frontal and with no shadows or reflections coming from lamps or other objects.
Art Recognition does not provide valuation services for artworks. We recommend reaching out to a professional art appraiser for assistance in determining the value of your art piece.