Raphael’s masterpiece may not be all his
One of the most debated faces in Renaissance art may not have been painted by the master himself. A new AI-driven study reveals that the face of St Joseph in Raphael’s Madonna della Rosa was likely created by someone else. Researchers applied a machine learning model trained on the artist’s known works and found that this particular face does not match Raphael’s signature style.
This finding supports a long-standing theory among art historians. Although the painting is often attributed to Raphael, questions have persisted for over a century regarding its authenticity. The application of AI now adds quantitative weight to those suspicions and highlights how modern tools can uncover new layers of insight within classical works.
What makes this face differentThe Madonna della Rosa, believed to have been completed between 1518 and 1520, has puzzled scholars for generations. While most of the painting displays the hallmarks of Raphael’s style, the face of St Joseph in the upper left corner appears weaker and less refined. Art critics have long suggested that this portion might have been completed by someone else in Raphael’s workshop.
The researchers used a two-step approach to examine this theory. First, the entire painting was evaluated using a deep learning classifier. Then, the image was divided into individual character sections and re-evaluated. The results showed that the Madonna, Child, and St John are highly consistent with Raphael’s style. However, the face of St Joseph was flagged as inconsistent.
How AI learned to see like RaphaelTo train the AI, the team used a transfer learning approach based on the ResNet50 convolutional neural network. This architecture was originally designed for general image recognition but was fine-tuned to detect the specific characteristics of Raphael’s work. Features such as brushstroke flow, shading, and edge patterns were extracted from authenticated paintings and used to train a support vector machine for classification.
In parallel, the researchers incorporated traditional edge detection methods, including Canny, Sobel, Laplacian of Gaussian, and Scharr operators. These techniques allowed the system to examine micro-level features that are difficult for the human eye to assess. The combination of deep features and edge maps produced a robust framework for style attribution.
One of the main challenges in applying machine learning to art history is the limited availability of training data. There are only a small number of authenticated Raphael paintings available for use. To address this issue, the researchers augmented the training set using image transformations such as rotations, translations, and mirroring. This expanded the effective dataset and helped the model generalize to new inputs.
The final model achieved 98 percent accuracy in validation tests. The results were compelling enough to apply the same method to several disputed or unknown works, including the Madonna della Rosa. When the face of St Joseph was isolated and analyzed, the classifier found a probability of only 37 percent that it had been painted by Raphael.
(Image credit)
The result supports a plausible explanation based on historical context. Renaissance artists often worked in collaborative studio environments, where assistants and pupils contributed to major works. Raphael’s workshop included students like Giulio Romano, who may have painted less prominent elements within a composition. While the exact artist behind St Joseph remains unknown, the AI findings lend credibility to the theory of shared authorship.
The team emphasizes that this discovery does not reduce the artistic value of the piece. Instead, it brings new understanding to the collaborative process behind many Renaissance artworks and offers a clear demonstration of how computational analysis can assist traditional connoisseurship.
To further test the model’s reliability, the researchers analyzed several known imitations and ambiguous paintings. One experiment involved a modern replica painted by one of the study’s authors. Although visually similar to Raphael’s work, the AI correctly identified it as an imitation. Another case involved a painting initially attributed to another artist but later suspected to be by Raphael. The model supported this reassessment, assigning it a high probability of being genuine.
These examples demonstrate that the model does more than detect superficial similarities. It captures the deeper, consistent elements that define Raphael’s style, including unique combinations of edge structure, tonal balance, and compositional flow.
One of the most innovative aspects of the study was the decision to analyze paintings in parts rather than only as wholes. This approach reflects how many Renaissance artists operated, with assistants often painting backgrounds or secondary figures. By assessing each part individually, the system provides more granular insight into authorship and stylistic consistency.
In the case of the Madonna della Rosa, most sections of the painting scored between 79 and 93 percent in favor of Raphael’s authorship. The face of Joseph was the exception. This anomaly gave researchers an opportunity to validate historical suspicions using computational evidence. It also illustrates how AI can detect stylistic discontinuities that are not immediately obvious to the naked eye.
Research: The gold standard for GenAI evaluation
Although this study focused on Raphael, the methodology is highly adaptable. Given a sufficient number of authenticated images, a similar model could be trained to evaluate the works of other historical artists. This approach could prove especially useful in cases where documentation is scarce or traditional forensic techniques provide inconclusive results.
The research team has already explored using higher-resolution image sets and other deep learning architectures. Early tests suggest that models like ResNet101 or EfficientNet may offer performance benefits for more complex tasks. However, the balance between computational cost and accuracy remains an open area for future exploration.
This study does not propose replacing experts but rather enhancing their ability to make informed judgments. AI models can process high-dimensional visual data more quickly and systematically than humans, making them valuable partners in verification, restoration, and cataloging efforts.
As machine learning tools continue to evolve, they are likely to become standard elements in the toolkit of art historians and curators. When integrated with provenance studies, pigment analysis, and iconographic research, AI can add a new dimension to our understanding of artistic heritage.
One face, many questionsThe face of St Joseph in the Madonna della Rosa has become a focal point for a larger discussion about authorship, collaboration, and the role of AI in the humanities. Thanks to this study, we now have a clearer picture of how machines can help answer questions that have puzzled scholars for generations.