Art, data and illusion · 2016
The Next Rembrandt: how artificial intelligence created a "new Rembrandt"
A man who never existed, a face calculated from hundreds of paintings and a surface printed in relief: the experience was spectacular. But who really created the image - the algorithm, the engineers or the history of art?
Immediate response
This is not a found Rembrandt, but a statistical portrait
The Next Rembrandt is a new image designed by a human team based on the numerical analysis of paintings attributed to Rembrandt. Algorithms measured faces, identified compositional regularities, synthesized a physiognomy and generated a texture. The result was then printed in several layers with ultraviolet sensitive ink to evoke the relief of a painting in oil.
The project was commissioned by ING and developed with the J. Walter Thompson Amsterdam agency, Microsoft, advisors from TU Delft, Mauritshuis and Museum Rembrandthuis. It was unveiled in Amsterdam on April 5, 2016, three hundred forty-seven years after the painter's death. His ambition was not to deceive a museum: it was a technological and cultural campaign designed to provoke a discussion about creativity.
Why this specific man?
The "typical topic" was first chosen by the team
The system was not given the vague order to "paint like Rembrandt". Portraits were classified into categories: male or female, age, facial orientation, hair, costume and accessories. The data indicated that a certain type appeared frequently: a middle-aged white man, bearded, looking to the right, dressed in black, wearing a white collar and hat.
This step already shows the decisive place of humans. Designers determine the works included, the useful categories and the profile to be produced. A statistic alone does not decide that a bearded man is more "rembranesque" than one elderly woman, child, biblical figure or self-portrait. She answers the question she was asked.
The final face is not that of any historical model. It results from a combination of medium shapes and geometric relationships derived from portraits. The presence effect comes precisely from this ambiguity: the character seems familiar, but no biography is behind his eyes.
An image base is never neutral
The project announced the analysis of 346 paintings. However, Rembrandt's catalog has changed a lot: certain works formerly autographs are today given to the workshop, to a student or to a follower. The chosen corpus therefore transmits its own uncertainties to the calculated image.
Not all files were homogeneous. They came from high definition scans, digital reproductions and other sources of varying quality. Before any generation, images had to be enlarged, aligned and standardized. In other words, the algorithm did not directly observe the paintings in museums: it worked on digital translations, already affected by lighting, calibration and resolution.
The process in six operations
Bring the corpus together
Partners assemble digital files of paintings attributed to Rembrandt and prepare them to become comparable.
Classify portraits
Works are annotated by gender, age, facial orientation, costume, beard, collar and hat.
Measure faces
A recognition system identifies more than sixty points and calculates the distances between eyes, nose, mouth, ears and facial contour.
Generate a physiognomy
The data makes it possible to construct a coherent face, a pose and a composition respecting the regularities of the chosen subset.
Simulate chiaroscuro and matter
Colors, light directions and height maps are calculated from reproductions and relief scans.
Print the surface
The image of more than 148 million pixels is printed in successive layers in order to reproduce a topography close to keys of painting.
AI didn't work alone in an empty room
The shortcut "artificial intelligence painted a Rembrandt" erases the chain of decisions. Humans defined the problem, selected the data, designed the categories, adjusted the results, determined the composition and chose the final version. The machines made analysis possible at unprecedented scale and speed, but they neither ordered the project nor formulated its intention.
It would also be anachronistic to confuse this 2016 experiment with the mainstream image generators that appeared later. The system was specialized, built for a single task and fed a delimited corpus. He does not did not respond to free sentences and did not instantly produce infinite variations.
What does he actually reproduce of Rembrandt's style?
Visible, immediately recognizable signs
The portrait takes on a familiar grammar: dark brown background, three-quarter bust, black suit, white collar, beret, side-lit face and slightly off-center gaze. The skin combines warm browns, redness and golden lights. Of small asymmetries avoid the overly geometric effect of a perfectly average face.
The printed surface adds a tactile argument. A flat image would have appeared to be a photomontage; the relief captures the light and evokes impasto. This materiality nevertheless remains a simulation: there is neither drying oil nor canvas worked by the brush, nor succession of decisions taken in front of one painting wet.
The algorithm excels at condensing what repeats itself. However, a great painter also recognizes himself in what he decides, at a specific moment, not to repeat.
Comparison: Rembrandt, workshop and calculated portrait
| Criterion | painting autograph | Workshop work | The Next Rembrandt |
|---|---|---|---|
| Author | Rembrandt, according to the state of knowledge | Student or collaborator in the master's environment | Contemporary team assisted by algorithms |
| Date | 17th century | 17th century, generally during the lifetime of the master | 2016 |
| Material | oil on canvas or panel | oil on canvas or panel | Digital image printed in layers with UV ink |
| Process | Pictorial decisions being executed | Learning, imitation, variation and sometimes intervention of the master | Analysis, generation, human selection and 3D printing |
| Status | Historical work | Historical work related to the workshop | Technological and cultural experience |
| Correct presentation | "Rembrandt" if attribution is accepted | "Rembrandt’s workshop" or museum formula | "Project The Next Rembrandt", never " painting from Rembrandt" |
Style is not a fixed recipe
Rembrandt changed his technique over the course of his career and depending on the function of the work. A social portrait from 1632, one tronie youthful and a late self-portrait present neither the same finish nor the same relationship to the model. By reducing the corpus to a dominant profile, the project obtains a strong general resemblance but attenuates this diversity.
The face of The Next Rembrandt seems convincing from a distance. On examination, some passages may seem too regular or illustrative. This is not a failure: the shift reveals what the data easily captures - composition, proportions, palette - and what resists more - intention, hesitation, accident and living relationship with the model.
Established facts and interpretations
Done
The portrait was unveiled in 2016 after approximately eighteen months of development.
Done
It is made up of more than 148 million pixels and is based on approximately 150 GB of graphics data.
Done
The character does not correspond to any known historical individual and the object is a relief print.
Interpretation
To say that AI is "artistic" depends on the definition chosen; the project does not philosophically resolve this question.
Artwork, technical demonstration or advertising campaign?
All three readings are possible and are not mutually exclusive. The object has a material presence, composition and ability to provoke emotion. It also demonstrates image processing, distributed computing and printing techniques relief. Finally, it was created for ING by an agency and became an extremely award-winning international campaign.
Caution is not to turn media success into proof of artistic authenticity. The project did not resurrect Rembrandt's consciousness. He constructed a statistical mirror of certain works that we associate with his name - mirror whose human choices are visible as soon as one examines the method.
How to look at the image for yourself?
Identify conventions
Spot the dark background, white collar, hat and side lighting. These are common signs, not a sufficient personal signature.
Follow the light
Ask yourself if it constructs a psychology and a space, or if it mainly reproduces a general appearance.
Examine transitions
Look at the passages between skin, hair, collar and bottom. The connecting zones often reveal the logic of synthesis.
Compare asymmetries
A living face is never perfectly average. Observe the difference between the two eyes and both sides of the mouth.
Separate image and object
On screen, the relief disappears. Faced with printing, real light accentuates the layers and modifies perception.
Reread the cartel
Check the collective author, date and technique. The vocabulary prevents us from confusing digital experience and 17th century work.
Where to see and understand the project?
Microsoft's official website maintains a detailed account, images and key technical figures. VML, heir to the agency that designed the campaign, presents the project in its archives. The object belongs to the heritage of the campaign rather than the permanent collection of a Rembrandt museum; its public presentations may therefore vary.
To compare with historical works, the Rijksmuseum in Amsterdam, the Mauritshuis in The Hague and the Museum Rembrandthuis remain the essential places. The best exercise is to observe the digital object first, then several authenticated portraits of different dates: the comparison reveals everything that an "average style" leaves aside.
See historical works
Explore Rembrandt without confusing original, workshop and generated image
The collection Alpha Reproduction clearly indicates known attributions. The Next Rembrandt not offered as painting from the master; to extend the analysis, explore the historical works and their notices.
Related articles
Frequently asked questions
Is The Next Rembrandt a real one painting from Rembrandt?
No. It is an object created in 2016 by a contemporary team using data from paintings by Rembrandt, then printed in relief.
Did artificial intelligence work alone?
No. Humans chose the corpus, categories, character type, generation method, final version and printing process.
How many works were analyzed?
The official project communication indicates 346 paintings and 168,263 image fragments.
Why is the character a bearded man in black?
Because the analysis and classification of the corpus led the team to select a frequent profile among the portraits studied: white man aged thirty to forty, with beard, white collar and hat.
Is it a painting oil?
No. The image was printed in several layers with UV ink to simulate the relief of a painted surface.
Can we talk about deep learning?
Partners use this expression for image processing and enlargement methods. The system was specialized and very different from current text-to-image generators.
Who owns the image rights?
This is a contemporary project linked to ING and the creative agency. It should not be treated as a work by Rembrandt that has fallen into the public domain.
Does the result prove that an AI can be an artist?
No, it opens the discussion. Its production combines automatic calculation, expertise, artistic direction, human selection and material manufacturing.
Where is The Next Rembrandt located?
The object belongs to the history of the countryside and can be presented in temporary exhibitions. It is not part of Rembrandt's painted catalog nor a permanent room dedicated to the master.
Main sources
- Microsoft Europe - The Next Rembrandt: Recreating the work of a master with AI.
- VML - official campaign archive The Next Rembrandt.
- TU Delft - context on The Next Rembrandt and robotic manufacturing.
- The Rembrandt Database - catalogs and questions of attribution of the painted corpus.
- Museum Rembrandthuis - scientific portal dedicated to Rembrandt.
The technical figures follow the publications of the project partners. The word "AI" here covers a set of specialized methods from 2016; the article deliberately distinguishes these methods from contemporary image generators.
0 comments