Afifi, Kalimullin, Korchagin and Kudryashov (arXiv:2511.22667, 2025) trained a tile-based convolutional neural network ensemble to recognize a single, static 'Rubens style' on a corpus of 141 works (85 Rubens, 56 stylistically comparable negatives) and applied it to disputed paintings. Their conclusions explicitly identify as an unresolved development direction that the system does not yet account for the evolution of the artist's style across periods and workshop phases, relying on expert-documented chronological information. The open problem is the construction of a temporally stratified, expert-verified corpus of the Rubens workshop oeuvre and of attribution algorithms that condition on this chronological and workshop structure, so that genuine stylistic evolution and workshop participation can be separated from misattribution.

Background

Computational attribution of Old Master paintings asks whether machine analysis of images can support — not replace — connoisseurship, the traditional art-historical practice of judging authorship from close visual comparison of securely attributed works. The paradigm case is Peter Paul Rubens, whose large workshop deliberately blurred the boundaries of authorship: pupils and collaborators executed major portions of canvases from the master's preparatory sketches, producing replicas, retouched workshop copies, and partial autographs that resist the binary categories demanded by the art market. The reference corpus for these debates is the Corpus Rubenianum Ludwig Burchard, the ongoing scholarly catalogue of Rubens's oeuvre. A recent preprint by Afifi, Kalimullin, Korchagin and Kudryashov (arXiv:2511.22667, 2025) contributes a deep-learning pipeline to this problem: an ensemble of convolutional neural networks (networks that learn visual features directly from image data) trained on small fixed-size image patches ('tiles') from high-resolution scans (about 5 pixels per millimeter) of 141 works — 85 by Rubens and 56 stylistically comparable negatives, including works by van Dyck and Jordaens — to recognize a single, static 'Rubens style', reaching approximately 85% classification accuracy on held-out test data, and applied to disputed paintings. Yet computational treatments of time remain at the level of movements and datasets, not of authentication: Elgammal, Liu, Kim, Elhoseiny and Mazzone (AAAI 2018) showed that CNN representations over 77,000 paintings reproduce the canonical art-historical chronology of styles, and the OmniArt multi-task framework of Strezoski and Worring (2017) predicts a work's creation year alongside style and artist — but neither conditions an authentication decision on a single artist's documented periods or workshop phases. Authentication studies themselves treat the artist's signature as static: Vision Transformers (an attention-based neural architecture that classifies images by treating image patches as a token sequence) for Van Gogh (Schaerf, Popovici and Postma, 2023), identification of the forger Beltracchi with Kolmogorov–Arnold networks (a neural architecture built from learnable one-dimensional functions) and CNNs (Boccuzzo, Meyer and Schaerf, 2024), machine learning on painted-surface topography (Ji et al., 2021), and one-class autoencoders (networks trained to reconstruct authentic works only, so that poor reconstruction signals probable non-authenticity) on handcrafted features for historical drawings (Ugail, Ritch-Frel and Matuzava, 2026) all learn one style model per artist. Related source-flagged directions have their own adjacent literatures but no solution: multimodal authentication combining recto (the painting's front) imaging with X-rays has been demonstrated for Goya (Ugail and Jaleel, 2025, using one-class classification — models trained on the authentic class only — on joint visual and radiographic features), and deep learning on macro X-ray fluorescence (MA-XRF) scanning — a non-invasive technique producing chemical element maps of a painting's sub-surface layers — is active in heritage-science processing (Yan et al., 2022; Verinaz-Jadan et al., 2024), while forgery benchmarks target generative-AI fakes (Aboutalebi et al., 2023) rather than historical pasticcio (a forgery assembled from fragments of several authentic works). What no found work does is model how a workshop artist's style changes over a career within the attribution decision itself, for Rubens or for any comparable workshop corpus.

Problem Statement

The problem is to construct a temporally stratified, expert-verified corpus of the Rubens workshop oeuvre — autograph works organized by documented period and workshop phase, documented collaborative and workshop productions, and stylistically comparable negatives — and to develop attribution algorithms that condition on this chronological and workshop structure rather than on a single static style model, so that genuine stylistic evolution across the artist's career and genuine workshop participation can be distinguished from misattribution. 'Misattribution' here denotes a work that is not autograph Rubens — a documented collaborative or workshop production, a copy, or a work by another hand — but is recorded, or is classified by the method, as autograph Rubens; documented workshop participation and stylistic evolution are the genuine phenomena to be preserved, so the documented collaborative productions and the stylistically comparable negatives are legitimate corpus classes rather than misattributions. The problem is to produce (i) a work-level stratification of the corpus that assigns each work to the documented period, and where applicable workshop phase, under which the expert catalogue literature places it, citing that documentation for each assignment, and (ii) a chronology- and workshop-conditioned attribution method evaluated on a held-out split against a single-static-style baseline trained on the same corpus; the scientific question is whether the conditioned method separates autograph works from the non-autograph classes better than the baseline, and a reported and analysed failure to do so is itself a complete resolution. Assembling this corpus and its period- and workshop-phase documentation from the scholarly catalogue literature — the Corpus Rubenianum Ludwig Burchard and comparable expert publications — is part of the solver's work rather than a supplied input. In the source's own terms, the system should be trained 'to take into account not only the artist's signature style but also its evolution depending on the period, workshop, ...', relying on 'already documented and verified information collected by experts' (Afifi et al., arXiv:2511.22667). The source bundles this with other development vectors (whole-image compositional analysis, verso (reverse-side) and X-ray multimodality, material-period pigment signals, pasticcio detection, protocols for expert-model interaction); those are related open directions in the same paper but do not share a completion boundary with this problem and are not required deliverables of it.

References

  1. Afifi, Angie; Kalimullin, Artur; Korchagin, Sergey; Kudryashov, Ilya. 'A deep learning perspective on Rubens' attribution'. arXiv:2511.22667 (2025). https://arxiv.org/abs/2511.22667
  2. Elgammal, Ahmed; Liu, Bingchen; Kim, Diana; Elhoseiny, Mohamed; Mazzone, Marian. 'The Shape of Art History in the Eyes of the Machine'. Proceedings of the AAAI Conference on Artificial Intelligence 32(1), 2018, DOI 10.1609/aaai.v32i1.11894. https://doi.org/10.1609/aaai.v32i1.11894
  3. Strezoski, Gjorgji; Worring, Marcel. 'OmniArt: Multi-task Deep Learning for Artistic Data Analysis'. arXiv:1708.00684 (2017). https://arxiv.org/abs/1708.00684
  4. Schaerf, Ludovica; Popovici, Carina; Postma, Eric. 'Art Authentication with Vision Transformers'. arXiv:2307.03039 (2023). https://arxiv.org/abs/2307.03039
  5. Ugail, Hassan; Jaleel, Ismail Lujain. 'Integrating Visual and X-Ray Machine Learning Features in the Study of Paintings by Goya'. arXiv:2511.01000 (2025). https://arxiv.org/abs/2511.01000
  6. Ugail, Hassan; Ritch-Frel, Jan; Matuzava, Irina. 'Handcrafted Feature-Assisted One-Class Learning for Artist Authentication in Historical Drawings'. arXiv:2601.11627 (2026). https://arxiv.org/abs/2601.11627
  7. Ji, F.; McMaster, M. S.; Schwab, S.; Singh, G.; Smith, L. N.; Adhikari, S.; O'Dwyer, M.; Sayed, F.; Ingrisano, A.; Yoder, D.; Bolman, E. S.; Martin, I. T.; Hinczewski, M.; Singer, K. D. 'Discerning the painter's hand: machine learning on surface topography'. arXiv:2106.07134 (2021). https://arxiv.org/abs/2106.07134
  8. Boccuzzo, Sandro; Meyer, Deborah Desirée; Schaerf, Ludovica. 'Art Forgery Detection using Kolmogorov Arnold and Convolutional Neural Networks'. arXiv:2410.04866 (2024). https://arxiv.org/abs/2410.04866
  9. Yan, Su; Huang, Jun-Jie; Verinaz-Jadan, Herman; Daly, Nathan; Higgitt, Catherine; Dragotti, Pier Luigi. 'A Fast Automatic Method for Deconvoluting Macro X-ray Fluorescence Data Collected from Easel Paintings'. arXiv:2210.17496 (2022). https://arxiv.org/abs/2210.17496
  10. Verinaz-Jadan, Herman; Yan, Su; Higgitt, Catherine; Dragotti, Pier Luigi. 'Adversarial Deep-Unfolding Network for MA-XRF Super-Resolution on Old Master Paintings Using Minimal Training Data'. arXiv:2409.09483 (2024). https://arxiv.org/abs/2409.09483
  11. Aboutalebi, Hossein; Mao, Dayou; Fan, Rongqi; Xu, Carol; He, Chris; Wong, Alexander. 'DeepfakeArt Challenge: A Benchmark Dataset for Generative AI Art Forgery and Data Poisoning Detection'. arXiv:2306.01272 (2023). https://arxiv.org/abs/2306.01272