Attribution of Old Master paintings increasingly rests on technical evidence — materials, layer structure, and paint handling — yet many works have undergone substantial historical restoration interventions (overpaint, loss compensation, varnish replacement, structural repair, even transposition of the paint layer to a new support) that alter exactly the material carriers of that evidence. The problem is how the contribution of such interventions can be separated from the original artist's own technique and style, so that attribution judgments reflect the original hand rather than the restorer's. Conservation science can localize non-original material within a single painting, but no validated general account exists for how restoration-altered evidence should be weighted in attribution; the literature from 2011 through 2026 treats the question case by case, and the Salvator Mundi controversy — where a heavily restored panel's attribution to Leonardo remains contested — is its flagship live instance.

Background

Attribution — deciding who made a work of art — is a foundational task of art history. It traditionally rested on connoisseurship, the expert's trained judgment of style and technique, and it is now standard to integrate that judgment with technical art history, the analysis of a work's materials and physical structure using conservation-science methods. Attribution evidence is carried materially: the choice and preparation of the support (panel or canvas), the ground (the preparatory layer on which paint is applied), the underdrawing — the artist's preparatory drawing beneath the paint, revealed by infrared reflectography, an infrared imaging method that penetrates the paint layers — and above all the paint handling — brushwork, layering, and blending — that connoisseurs read as an artist's hand.

Restoration interventions interfere with exactly this evidence. Historic treatments include overpaint (later paint covering original passages), inpainting (compensation of paint losses), varnish removal and replacement, structural repairs, and in extreme cases transposition — transferring the entire paint layer to a new support, an operation that risks damaging the original ground and paint structure. Attribution judgments on such paintings must therefore work from evidence admixed with, and partly produced by, later hands.

Conservation science can localize much of this foreign material within a given painting. X-radiography reveals internal structure and losses; ultraviolet fluorescence exposes recent retouching; cross-sections (minute sampled stratigraphies of the paint buildup) and pigment-analysis methods such as Raman spectroscopy and polarized light microscopy identify materials; and optical coherence tomography (OCT, infrared interferometric imaging of subsurface layers) can visualize original versus non-original semi-transparent layers, as multi-scale OCT did during the 2016–21 Mauritshuis treatment of Rogier van der Weyden's Lamentation of Christ — a painting whose attribution and dating have been debated for at least two centuries. In other recent cases the same pattern holds: the Madone Lansdowne (La Vierge aux fuseaux), a panel twice transposed (1911 and 1970) whose attribution in relation to Leonardo remains discussed, was re-examined during its 2016–18 Opificio delle Pietre Dure restoration; and a painting long attributed to Salvator Rosa was de-attributed only through restoration and laboratory analysis, which suggested the false attribution arose from fraudulent documents at the time of sale. The Lansdowne panel also shows how restoration can irreversibly remove evidence: after its two transpositions, the original support can no longer be identified, and whether the surviving ground layer is original cannot be established.

What the field possesses is the detection half of a separation procedure, not the inferential half. Attribution reasoning on compromised paintings proceeds case by case: experts argue from the passages believed original, but there is no settled, transferable account of how the kind and extent of alteration should weight, or limit, the conclusions one may draw — how much attribution weight a given quantum of surviving original material can bear. The question became public with the Salvator Mundi, a Leonardo-attributed panel whose surface was substantially reconstructed during restoration by the conservator Dianne Modestini; scholarly opinion on its attribution remains split, and a convolutional-neural-network authorship test (a deep-learning image classifier trained to recognize artist-specific visual features) explicitly identified the extent of damage and restoration as a confounding factor it could not separate from authorship signal, concluding that Leonardo's few extant paintings test the limits of the system. Automated brushstroke analysis has shown that quantitative technique signatures can distinguish hands on well-preserved works (for example, separating van Gogh from contemporaries), and patch-based neural-network classifiers have been applied to single disputed paintings, but each of these efforts defers the restoration problem rather than solving it.

Problem Statement

When a painting has undergone substantial historical restoration interventions that altered its original material structure — overpaint, inpainting of losses, varnish replacement, structural repair, or transposition of the paint layer to a new support — how can the contribution of those interventions be separated from the original artist's own technique and style, so that attribution judgments, including their stated uncertainty, rest on original material rather than on evidence introduced or transformed by the restorer?

The state of the art localizes non-original material within a given painting by scientific imaging and analysis. The open core is the general inferential account: a transferable framework for how attribution evidence should be weighted as a function of the type, extent, and detectability of the alteration, applicable across cases rather than constructed ad hoc for each painting, together with an argument grounded in the cited cases for why it applies across them. The answer sought is such a framework that is general in the sense just stated, applicable across the cited cases rather than to a single painting, and that sets out how the weighted evidence issues in an attribution judgment together with that judgment's stated uncertainty.

References

  1. Cinzia Pasquali, "La Vierge aux fuseaux – Madone Lansdowne. Hypothèses sur l'influence de Léonard de Vinci dans l'exécution de l'œuvre", Technè, Hors-série, 2024, pp. 116–127. DOI: 10.4000/11v47, https://doi.org/10.4000/11v47
  2. Carol Pottasch, Abbie Vandivere, Lieve d'Hont, "Unveiling the Lamentation of Christ by Rogier van der Weyden and his workshop: Discoveries made during a restoration treatment", Bulletin de l'Institut royal du Patrimoine artistique, vol. 41, 2025. DOI: 10.4000/14m38, https://doi.org/10.4000/14m38
  3. Mariana Calderón Mejía, Marcos Tascon, Damasia Gallegos, Dolores Gónzalez Pondal, Mariana Bini, Lucia Liccioli, Serena Barone, Lorenzo Giuntini, Marielena Fedi, Francesco Taccetti, Fernando Marte, "The role of restoration and scientific examination for the accurate attribution of a European painting in South America", Rendiconti Lincei. Scienze Fisiche e Naturali, vol. 31, no. 2, 2020, pp. 431–442. DOI: 10.1007/s12210-020-00904-y, https://doi.org/10.1007/s12210-020-00904-y
  4. Martin Kemp, "Art history: Sight and salvation", Nature, vol. 479, 2011, pp. 174–175. DOI: 10.1038/479174a, https://doi.org/10.1038/479174a
  5. Steven J. Frank, Andrea M. Frank, "A Neural Network Looks at Leonardo's(?) Salvator Mundi", arXiv:2005.10600, 2020. https://arxiv.org/abs/2005.10600
  6. Jean-Pierre Crettez, "Internal Geometry of "Salvator Mundi" (The "Cook Version", Attributed to Leonardo da Vinci)", in Mathematics in the Visual Arts, Wiley-ISTE, 2020, pp. 103–126. DOI: 10.1002/9781119801801.ch6, https://doi.org/10.1002/9781119801801.ch6
  7. Jia Li, Lei Yao, E. Hendriks, J. Z. Wang, "Rhythmic Brushstrokes Distinguish van Gogh from His Contemporaries: Findings via Automated Brushstroke Extraction", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 6, 2012, pp. 1159–1176. DOI: 10.1109/TPAMI.2011.203, https://doi.org/10.1109/TPAMI.2011.203
  8. Juan Ruiz de Miras, José Luis Vílchez, María José Gacto, Domingo Martín, "Painting authentication using CNNs and sliding window feature extraction", Frontiers in Artificial Intelligence, vol. 8, article 1738444, 2026. DOI: 10.3389/frai.2025.1738444, https://doi.org/10.3389/frai.2025.1738444
  9. Nobuyuki Kamba, "Multimodal Non-Invasive Investigation of the Penitent Magdalene, Formerly in the Joseph Robinson Collection: New Technical Evidence for Attribution", Heritage, vol. 9, no. 8, article 324, 2026. DOI: 10.3390/heritage9080324, https://doi.org/10.3390/heritage9080324