Pith. sign in

REVIEW 4 major objections 5 minor 58 references

PyPotteryInk: One-Step Diffusion Model for Sketch to Publication-ready Archaeological Drawings

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PyPotteryInk claims a one-step diffusion model converts pencil sketches of archaeological pottery into publication-ready inked drawings in seconds while preserving the vessel profiles, decorations, and structural details that experts rely…

desk verdict A genuinely useful open tool for automated inking of pottery drawings, but its 'publication-ready' claim rests on a small, statistically unanalyzed expert study. read the letter →

arxiv 2502.06897 v1 pith:C7MMJRFT submitted 2025-02-09 cs.GR cs.AIcs.CV

Lorenzo Cardarelli This is my paper · ORCID
classification cs.GRcs.AIcs.CV
keywords PotteryGenerativeAIArchaeologicalDrawingImage-to-ImageTranslationDiffusionModelsOne-stepPatch-basedinferenceFine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the slow, skilled step of inking archaeological pottery drawings by hand can be automated without losing the information the drawings exist to carry. It introduces PyPotteryInk, a one-step diffusion model trained on pairs of pencil sketches and their inked versions from Italian protohistoric pottery. The authors argue that the model preserves morphological and decorative details, produces high-resolution output of any drawing size through patch-based processing, and can be fine-tuned with only a handful of examples for other pottery styles. If true, this would cut a documentation step that takes hours down to seconds and make more assemblages publishable with the same professional effort. The paper's central claim is that expert evaluation confirms the outputs meet archaeological publication standards.

What carries the argument

The load-bearing mechanism is the img2img-turbo architecture, a one-step image-to-image translation variant of Stable Diffusion. The input sketch is encoded into a latent space, processed by a modified U-Net in one denoising pass, and decoded back into image space; a fixed text prompt keeps the output style consistent. The model is trained with reconstruction, perceptual, adversarial, and CLIP-alignment losses, with LoRA low-rank adapters so that fine-tuning to a new pottery style or documentation convention needs only a small parameter update. To keep resolution high, inference splits the drawing into 512-by-512 patches with a 64-pixel overlap and recombines the processed patches, which is what lets arbitrarily large drawings be handled without losing fine details.

What would settle it

A controlled test in which independent archaeologists, blind to the condition, rate the same model outputs against traditionally inked versions of the same sketches using a pre-registered rubric, and also decide whether each drawing is ready for publication, would settle the claim; if model outputs are rated below the human-inked versions or are reliably distinguished by experts above chance, the 'meets publication standards' claim would be falsified.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that PyPotteryInk generates inked drawings of consistently high quality that meet archaeological standards. The model takes a pencil sketch as input and, in a single forward pass of a modified latent diffusion network, outputs a clean inked drawing with homogeneous lines, stippled shading, and the standard profile-and-prospectus layout. It preserves structural elements such as rims, handles, fractures, and profile sections as well as fine decorative patterns, because the patch-based inference with dynamic overlap lets the network work on the full original resolution rather than a downscaled square. Expert assessment, including a single-blind discrimination test in which most evaluators could not reliably tell machine-inked from hand-inked drawings, is presented as evidence that the output meets the analytical requirements of the field. The author frames the tool as complementary rather than a replacement: a human still produces the original sketch, and the output can be manually edited.

Load-bearing premise

The claim that outputs are publication-ready rests on a single study with four expert evaluators whose averaged ratings and a blind discrimination test are treated as a valid measure of archaeological publication quality.

Editorial extensions

If this is right

  • Archaeological illustrators could shift time from manual inking to interpretation and editing, since the model produces a first-pass ink drawing in seconds.
  • Unpublished or backlogged pottery assemblages could be documented faster, increasing the volume of comparable published drawings among assemblages.
  • A small fine-tuning set of roughly ten to twenty paired drawings could adapt the model to a different drawing tradition or material class, if the style differences are not too extreme.
  • The fixed output style means all drawings from a context could be standardized, making morphological comparison between assemblages more consistent.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next test, which the paper does not carry out, is a direct head-to-head comparison against manually inked drawings judged by a larger independent panel of archaeologists, rather than a comparison with the original pencil sketch.
  • Because the model is open-source and fine-tunes on few examples, it could plausibly extend to lithics, metal objects, or other documentation traditions, though the paper only demonstrates Italian protohistoric pottery.
  • The model's dependence on GPU hardware is a practical barrier for many archaeology teams; a distilled or quantized variant, or more accessible cloud workflows, would lower the entry cost for routine use.
  • The reported confusion between decoration and shading on styles outside the training distribution suggests that fine-tuning on painted wares would be needed before the method generalizes beyond the impasto tradition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents PyPotteryInk, an open-source pipeline based on a one-step diffusion model (a modified img2img-turbo) that converts pencil sketches of archaeological pottery into inked, publication-ready drawings. The model is trained on 492 paired drawings from Italian protohistoric contexts, fine-tuned with heavy augmentation on a small Monte Croce Guardia set, and evaluated on 72 held-out Montale sketches. Validation consists of training/validation loss curves, a single-blind discrimination test in which four archaeological experts classify drawings as AI-generated or traditional, and qualitative 1–5 ratings of six criteria by a different set of four experts. The paper's central assertion in §5.1 is that the outputs meet archaeological publication standards and are produced in seconds rather than hours.

Significance. If the central claim were fully supported, the tool would be practically valuable, reducing a well-known bottleneck in archaeological ceramic publication and enabling broader dissemination of assemblages. The paper has real strengths: the code, pre-trained models, and documentation are stated to be publicly available; the 72-image Montale test set is genuinely held out; expert evaluators are named and represent domain practitioners; and the limitations section candidly discusses style-transfer failures and hardware requirements. However, the evidence currently offered does not establish 'publication-ready' or 'seconds rather than hours' with the rigor those claims require. The expert evaluation is small, under-analyzed, and lacks a human-inked baseline, so while the contribution is promising, its significance is not yet demonstrated.

major comments (4)
  1. [§4.2.1.1, Table 1] The single-blind discrimination test is reported only as aggregate accuracy/precision/recall/Jaccard per expert, without the number of items, per-drawing classifications, confidence intervals, or a binomial error model. This makes the results load-bearing but uninterpretable: EP1's accuracy of 0.05 is explained via personal communication as a preference for clean AI outputs, yet with no stated n or catch trials it is equally consistent with the expert misunderstanding the task or the protocol failing. At the same time, FE's 0.85 accuracy with precision 1.0 shows that at least one expert could reliably distinguish AI-generated from traditional drawings, so the claim in §5.1 that 'most evaluators had difficulty' is only partially supported and is contradicted for FE. Please report per-item data, add catch or control drawings, and provide error bars or binomial tests for each expert.
  2. [§4.2.1.2, Table 2] The qualitative evaluation reports only mean scores (1–5) for six criteria, with no variance, per-drawing distributions, or inter-rater reliability statistics, and the four raters in Table 2 are not the same as those in Table 1. More importantly, no human-inked drawings were rated under the same criteria, so the averages do not establish non-inferiority to the accepted product. The 'publication-ready' claim is further weakened by the FD criterion, which directly asks whether the drawing is complete without manual additions: CP's mean is 2.74 (between 'poor' and 'satisfactory') while ADR's is 4.98, a disagreement that the current reporting hides. Please report the full score distributions, add a human-inked control set rated under identical criteria, and compute inter-rater reliability (e.g., ICC or Fleiss' kappa).
  3. [§5.1 / Abstract] The claim that the model produces drawings 'in seconds rather than the hours required for manual inking' is not supported by any measurement reported in the manuscript; it rests on comments attributed to CP and EP2. Since this is one of the two headline benefits in the abstract and Section 2, please report measured per-patch and per-drawing inference times on the stated hardware (RTX 3070Ti), and either compare them with documented manual inking times or explicitly label the speed claim as anecdotal.
  4. [§4.1 / §4.2] The only quantitative results are training and validation curves; no held-out quantitative evaluation (e.g., FID or LPIPS on the 72 Montale test images) is reported. This absence is not fatal by itself, but it means the entire 'publication-ready' claim rests on the expert study, which makes the reporting deficiencies in the previous comments decisive. Adding test-set metrics would provide an independent, reproducible check on the model's generalization.
minor comments (5)
  1. [Table 2] In Table 2, the expert row labeled 'AC' collides with the criterion column 'AC' (archaeological consistency), and the value '4,98' in the ADR row uses a comma as decimal separator; please rename the row and unify decimal notation.
  2. [§4.2.1.1, footnotes 3–5] The formal definitions of recall, precision, and Jaccard are worded incorrectly; for example, recall should be 'the proportion of AI-generated drawings correctly identified as AI among all actual AI-generated drawings', not 'among the total of correct identifications'.
  3. [§3.2, Eq. (1)] The symbol x is used both for the input sketch in the encoding step and for the target inked drawing in the reconstruction loss L_rec; please use distinct notation (e.g., x_in and y).
  4. [§3.4] The inference procedure list does not mention the 64-pixel overlap discussed in the preceding paragraph; please incorporate it into the step list or clarify how the overlap is applied during recombination.
  5. [§3.3–§3.4] Key training hyperparameters (learning rate, batch size, LoRA rank, and loss weights) are not reported in the text, only deferred to the repository; please include the most important ones in the paper for self-contained reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is trained on paired external drawings and evaluated by independent expert judgments, so the reported predictions do not reduce to fitted inputs or self-citations.

full rationale

PyPotteryInk's derivation chain is supervised image-to-image translation: paired pencil sketches and inked drawings from external archaeological sources are used to train a one-step diffusion model, and the output quality is assessed on held-out sketches by independent expert raters and a single-blind discrimination test. None of the reported metrics—LPIPS, clean-FID, expert AC/LQ/SQ/OQ/FR/FD scores, or discrimination accuracy—is defined in terms of the model's own fitted parameters or outputs; they compare against ground-truth inked drawings or external human judgments. The only self-citations (Cardarelli 2022, 2023, 2024a, 2024b) appear as contextual examples of prior AI applications in archaeology and are not load-bearing for the central claim. The architecture and losses are attributed to external prior work (Parmar et al. 2024, Rombach et al. 2022, Hu et al. 2021), and no uniqueness theorem or ansatz is smuggled in through self-citation. The main scientific weaknesses—a small expert panel, absence of inter-rater reliability statistics, no human-inked baseline under the same rubric, and a possible ambiguity about Montale data appearing both in the training corpus and in the held-out test—are evidentiary and correctness concerns, not circularity: the evaluation remains an external benchmark rather than a quantity reconstructed from training data. Accordingly, no circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a standard paired image-to-image translation setup, a borrowed one-step diffusion architecture, a curated but withheld dataset, and subjective expert evaluation. No derivational free parameters beyond the listed hyperparameters are required, and no new physical or conceptual entities are introduced.

free parameters (6)
  • patch_size = 512x512 pixels
    Inference and fine-tuning are performed on 512x512 patches; the ability to process large drawings and preserve detail depends on this choice (Section 3.4).
  • patch_overlap = 64 pixels
    Dynamic overlap between patches is introduced to avoid splitting artifacts; the value is chosen by hand and no ablation is shown (Section 3.4).
  • training_steps_10k = 10,000 steps
    Duration of base model training; convergence is reported around step 8000, and the step count is chosen without a formal stopping criterion (Sections 3.3 and 4.1).
  • fine_tuning_steps_6h = 600 steps
    Fine-tuning duration on 9 Monte Croce Guardia pairs; no ablation on this number is provided (Section 3.4).
  • fixed_prompt = "make it ready for publication"
    A fixed text condition is used to enforce a consistent output style; its effect is not ablated (Section 3.2).
  • loss_weights = not specified in text, in repository
    The balance between reconstruction, adversarial, and CLIP losses determines output quality, but the values are not reported in the manuscript (Section 3.2).
assumptions (3)
  • domain assumption Training pairs are correctly aligned and representative of the sketch-to-ink mapping.
    Paired image translation assumes the pencil sketch and inked version correspond to the same vessel in a consistent spatial relation; the paper resizes pairs to 512x512 but never describes alignment verification (Section 3.3).
  • domain assumption Expert ratings and discrimination scores measure publication readiness.
    The single-blind test and 1-5 qualitative ratings are treated as ground truth for publication standards; no inter-rater reliability statistics or external benchmark is provided (Section 4.2.1).
  • domain assumption The one-step distilled U-Net from img2img-turbo is capable of preserving fine morphological details in this domain.
    The architecture is borrowed from prior work and assumed able to represent decorative patterns and vessel profiles at the chosen patch resolution; no ablations on this capability are shown (Sections 3.2 and 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PyPotteryInk: One-Step Diffusion Model for Sketch to Publication-ready Archaeological Drawings." pith.science (2026). https://pith.science/paper/C7MMJRFT

@misc{pith2026250206897,
  author       = {Pith},
  title        = {Pith review of: PyPotteryInk: One-Step Diffusion Model for Sketch to Publication-ready Archaeological Drawings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7MMJRFT}},
  note         = {Machine review of arXiv:2502.06897}
}
read the original abstract

Archaeological pottery documentation traditionally requires a time-consuming manual process of converting pencil sketches into publication-ready inked drawings. I present PyPotteryInk, an open-source automated pipeline that transforms archaeological pottery sketches into standardised publication-ready drawings using a one-step diffusion model. Built on a modified img2img-turbo architecture, the system processes drawings in a single forward pass while preserving crucial morphological details and maintaining archaeologic documentation standards and analytical value. The model employs an efficient patch-based approach with dynamic overlap, enabling high-resolution output regardless of input drawing size. I demonstrate the effectiveness of the approach on a dataset of Italian protohistoric pottery drawings, where it successfully captures both fine details like decorative patterns and structural elements like vessel profiles or handling elements. Expert evaluation confirms that the generated drawings meet publication standards while significantly reducing processing time from hours to seconds per drawing. The model can be fine-tuned to adapt to different archaeological contexts with minimal training data, making it versatile across various pottery documentation styles. The pre-trained models, the Python library and comprehensive documentation are provided to facilitate adoption within the archaeological research community.

Figures

Figures reproduced from arXiv: 2502.06897 by the authors.

Figure 1
Figure 1. Training dataset for the ‘10k’ model. Images are “compressed” into a squared box for training purpose [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Training dataset for the ‘6h-MCG’ model. Extensive data augmentation is used In this example, the original 1832 · 885 pixels image is divided into 8 patches (512 · 512 as standard size, plus additional non-standard patches to fill in gaps) that are processed one at a time, representing the native resolution of the fine-tuned ‘6h-MCG’ model. The inference algorithm thus ensures that drawings of any size can be proces… view at source ↗
Figure 3
Figure 3. Inference patching for an example image. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Training metrics for the 10k model [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Validation metrics for the 10k model 7 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: 10k model’s output during the training. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: 6h-MCG model’s output. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Some of the AI-inked drawings from Montale assemblage [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Limitations of the proposed model: given the completely different style, the model cannot reproduce shading [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 39 canonical work pages

  1. [1]

    Altaweel, Mark, Adel Khelifi, and Mohammad Hashir Zafar. 2024. ``Using Generative AI for Reconstructing Cultural Artifacts : Examples Using Roman Coins .'' Journal of Computer Applications in Archaeology 7 (1). https://doi.org/10.5334/jcaa.146

    work page doi:10.5334/jcaa.146 2024

  2. [2]

    Amanbay, Makhabbat. 2023. ``The Ethics of AI -Generated Art .'' \ SSRN\ \ Scholarly\ \ Paper\ . Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=4551467

    work page 2023

  3. [3]

    Anichini, Francesca, Nachum Dershowitz, Nevio Dubbini, Gabriele Gattiglia, Barak Itkin, and Lior Wolf. 2021. ``The Automatic Recognition of Ceramics from Only One Photo: The ArchAIDE App.'' Journal of Archaeological Science: Reports 36 (April): 102788. https://doi.org/10.1016/j.jasrep.2020.102788

    work page arXiv 2021

  4. [4]

    Azatbekova, Nurzada. 2023. ``Is Artificial Intelligence Reshaping Our World ? Exploring the Revolutionary Impact of AI in Everyday Life .'' European Research Materials, no. 3 (July). https://ojs.scipub.de/index.php/ERM/article/view/1898

    work page 2023

  5. [5]

    Cardarelli, I

    Barbaro, Barbara, A. Cardarelli, I. Damiani, F. di Gennaro, Nicola Ialongo, Andrea Schiappelli, and F. Trucco. 2011. ``Monte Cimino ( Soriano Nel Cimino , VT ) : Un Centro Fortificato e Un Complesso Cultuale Dell'età Del Bronzo Finale : Rapporto Preliminare.'' Scienze Dell'Antichità 17: 611--20. https://www.torrossa.com/it/resources/an/3087274

    work page arXiv 2011

  6. [6]

    Bickler, Simon H. 2021. ``Machine Learning Arrives in Archaeology .'' Advances in Archaeological Practice 9 (2): 186--91. https://doi.org/10.1017/aap.2021.6

    work page doi:10.1017/aap.2021.6 2021

  7. [7]

    Cacciari, I., and G. F. Pocobelli. 2022. ``Machine Learning : A Novel Tool for Archaeology .'' In Handbook of Cultural Heritage Analysis , edited by Sebastiano D'Amico and Valentina Venuti, 961--1002. Cham: Springer International Publishing. https://doi.org/10.1007/978-3-030-60016-7_33

    work page doi:10.1007/978-3-030-60016-7_33 2022

  8. [8]

    Cardarelli, Andrea. 2009. Guida Al Parco Archeologico e Museo All'aperto Della Terramare Di Montale . Digital Index Editore. https://digitalindex.it/guida-al-parco-archeologico-e-museo-all-aperto-della-terramare-di-montale

    work page 2009

Show all 58 references
  1. [9]

    ---------. 2014. La Necropoli Della Terramara Di Casinalbo . Grandi Contesti e Problemi Della Protostoria Italiana 15. Borgo San Lorenzo (Fi): All'insegna del giglio

    2014

  2. [10]

    ---------. 2018. ``Before the City. The Last Villages and Proto-Urban Centres Between the Po and Tiber Rivers.'' Origini 42 (2): 359--82. https://iris.uniroma1.it/handle/11573/1335640

    2018

  3. [11]

    Cardarelli, Andrea, Marco Bettelli, Andrea Di Renzoni, Maurizio Cruciani, and Nicola Ialongo. 2017. ``Nuove Ricerche Nell'abitato Della Tarda Età Del Bronzo Di Monte Croce Guardia ( Arcevia -- AN ) : Scavi 2015-2016.'' Rivista Di Scienze Preistoriche : LXVII, 2017, no. LXVII. ...

    2017 doi

  4. [12]

    Cardarelli, A., and F. Trucco. 2014. ``Monte Cimino : Abitato Fortificato e Centro Cerimoniale Di Sommità Nell' Etruria Protostorica Alle Soglie Della Svolta Protourbana.'' In Etruria in Progress , La Ricerca Archeologica in Etruria Meridionale , edited by L. Mercuri and R. Za...

    2014

  5. [13]

    Cardarelli, Lorenzo. 2022. ``A Deep Variational Convolutional Autoencoder for Unsupervised Features Extraction of Ceramic Profiles. A Case Study from Central Italy .'' Journal of Archaeological Science 144 (August): 105640. https://doi.org/10.1016/j.jas.2022.105640

    2022

  6. [14]

    ---------. 2023. ``Traditional and Digital Typologies Compared: The Example of Italian Protohistory.'' Origini XLVII. https://www.torrossa.com/it/catalog/preview/5751207

    2023

  7. [15]

    ---------. 2024a. Morphological Variability and Standardisation of Vessel Shapes in the 2nd and First Half of the First Millennium BC in Continental Italy . Adrias. IT: EDIPUGLIA SRL. https://doi.org/10.4475/0840

    doi

  8. [16]

    ---------. 2024b. ``From Fragments to Digital Wholeness: An AI Generative Approach to Reconstructing Archaeological Vessels.'' Journal of Cultural Heritage 70 (November): 250--58. https://doi.org/10.1016/j.culher.2024.09.012

    2024 doi

  9. [17]

    Roosevelt

    Demján, Peter, Peter Pavúk, and Christopher H. Roosevelt. 2023. ``Laser- Aided Profile Measurement and Cluster Analysis of Ceramic Shapes .'' Journal of Field Archaeology 48 (1): 1--18. https://doi.org/10.1080/00934690.2022.2128549

    2023

  10. [18]

    Dhariwal, Prafulla, and Alex Nichol. 2021. ``Diffusion Models Beat GANs on Image Synthesis .'' arXiv. https://doi.org/10.48550/arXiv.2105.05233

    2021 doi arXiv

  11. [19]

    Elliott, Anthony. 2019. The Culture of AI : Everyday Life and the Digital Revolution . 1st ed. Routledge. https://doi.org/10.4324/9781315387185

    2019 doi

  12. [20]

    Gattiglia, Gabriele. 2025. ``Managing Artificial Intelligence in Archeology . An Overview.'' Journal of Cultural Heritage 71 (January): 225--33. https://doi.org/10.1016/j.culher.2024.11.020

    2025 doi

  13. [21]

    Goodfellow, Ian J., Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. ``Generative Adversarial Networks .'' arXiv. https://doi.org/10.48550/arXiv.1406.2661

    2014 doi arXiv

  14. [22]

    Anne Jenner, and Christine Wilson

    Griffiths, Nick, M. Anne Jenner, and Christine Wilson. 2002. Drawing Archaeological Finds: A Handbook. Repr. Occasional Paper ... Of the Institute of Archaeology , University College London 13. London: Archetype Publ

    2002

  15. [23]

    Gui, Jie, Zhenan Sun, Yonggang Wen, Dacheng Tao, and Jieping Ye. 2020. ``A Review on Generative Adversarial Networks : Algorithms , Theory , and Applications .'' arXiv. https://doi.org/10.48550/arXiv.2001.06937

    2020 doi arXiv

  16. [24]

    Hu, Edward J., Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. `` LoRA : Low - Rank Adaptation of Large Language Models .'' arXiv. https://doi.org/10.48550/arXiv.2106.09685

    2021 doi arXiv

  17. [25]

    Hunt, Alice, ed. 2016. The Oxford Handbook of Archaeological Ceramic Analysis . 1st ed. Oxford University Press. https://doi.org/10.1093/oxfordhb/9780199681532.001.0001

    2016

  18. [26]

    Isola, Phillip, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. 2018. ``Image-to- Image Translation with Conditional Adversarial Networks .'' arXiv. https://doi.org/10.48550/arXiv.1611.07004

    2018 doi arXiv

  19. [27]

    Jaramillo, Pablo, and Ivan Sipiran. 2024. ``Cultural Heritage 3D Reconstruction with Diffusion Networks .'' arXiv. https://doi.org/10.48550/arXiv.2410.10927

    2024 doi arXiv

  20. [28]

    Kadhim, Israa, and Fanar M. Abed. 2023. ``A Critical Review of Remote Sensing Approaches and Deep Learning Techniques in Archaeology .'' Sensors 23 (6): 2918. https://doi.org/10.3390/s23062918

    2023 doi

  21. [29]

    Kingma, Diederik P., and Max Welling. 2022. ``Auto- Encoding Variational Bayes .'' arXiv. https://doi.org/10.48550/arXiv.1312.6114

    2022 doi arXiv

  22. [30]

    Le, Quan, Luis Miralles-Pechuán, Shridhar Kulkarni, Jing Su, and Oisín Boydell. 2020. ``An Overview of Deep Learning in Industry .'' In Data Analytics and AI , edited by Jay Liebowitz, 1st ed., 65--98. Auerbach Publications. https://doi.org/10.1201/9781003019855-5

    2020 doi

  23. [31]

    Levi, Sara T. 2010. Dal Coccio Al Vasaio: Manifattura, Tecnologia e Classificazione Della Ceramica. 1. ed. Bologna: Zanichelli

    2010

  24. [32]

    Ling, Ziyao, Giovanni Delnevo, Paola Salomoni, and Silvia Mirri. 2024. ``Findings on Machine Learning for Identification of Archaeological Ceramics : A Systematic Literature Review .'' IEEE Access 12: 100167--85. https://doi.org/10.1109/ACCESS.2024.3429623

    2024

  25. [33]

    Lyons, Mike, Franziska Fecher, and Markus Reindel. 2022. ``From LiDAR to Deep Learning: A Case Study of Computer-Assisted Approaches to the Archaeology of Guadalupe and Northeast Honduras .'' It - Information Technology 64 (6): 233--46. https://doi.org/10.1515/itit-2022-0004

    2022 doi

  26. [34]

    Navarro, Pablo, Celia Cintas, Manuel Lucena, José Manuel Fuertes, Claudio Delrieux, and Manuel Molinos. 2021. ``Learning Feature Representation of Iberian Ceramics with Automatic Classification Models.'' Journal of Cultural Heritage 48 (March): 65--73. https://doi.org/10.1016/...

    2021 doi

  27. [35]

    Navarro, Pablo, Celia Cintas, Manuel Lucena, José Manuel Fuertes, Rafael Segura, Claudio Delrieux, and Rolando González-José. 2022. ``Reconstruction of Iberian Ceramic Potteries Using Generative Adversarial Networks.'' Scientific Reports 12 (1): 10644. https://doi.org/10.1038/...

    2022 doi

  28. [36]

    Orton, Clive, and Michael Hughes. 2013. Pottery in Archaeology . 2nd ed. Cambridge University Press. https://doi.org/10.1017/CBO9780511920066

    2013 doi

  29. [37]

    Pang, Honglin, Xiujin Qi, Chengjun Xiao, Ziying Xu, Guangchen Ding, Yi Chang, Xi Yang, and Tianjing Duan. 2024. ``Pottery Evolution Pattern Discovery Based on Deep Learning: Case Study of Miaozigou Culture in China .'' Heritage Science 12 (1): 352. https://doi.org/10.1186/s404...

    2024 doi

  30. [38]

    Parisotto, Simone, Ninetta Leone, Carola-Bibiane Schönlieb, and Alessandro Launaro. 2022. ``Unsupervised Clustering of Roman Potsherds via Variational Autoencoders .'' Journal of Archaeological Science 142 (June): 105598. https://doi.org/10.1016/j.jas.2022.105598

    2022

  31. [39]

    Parmar, Gaurav, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. 2024. ``One- Step Image Translation with Text -to- Image Models .'' arXiv. https://doi.org/10.48550/arXiv.2403.12036

    2024 doi arXiv

  32. [40]

    Parmar, Gaurav, Richard Zhang, and Jun-Yan Zhu. 2022. ``On Aliased Resizing and Surprising Subtleties in GAN Evaluation .'' arXiv. https://doi.org/10.48550/arXiv.2104.11222

    2022 doi arXiv

  33. [41]

    Pawlowicz, Leszek M., and Christian E. Downum. 2021. ``Applications of Deep Learning to Decorated Ceramic Typology and Classification: A Case Study Using Tusayan White Ware from Northeast Arizona .'' Journal of Archaeological Science 130 (June): 105375. https://doi.org/10.1016...

    2021

  34. [42]

    Peroni, Renato. 1994. Introduzione Alla Protostoria Italiana. 1. Ed 47. Roma: Laterza

    1994

  35. [43]

    Radford, Alec, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, et al. 2021. ``Learning Transferable Visual Models From Natural Language Supervision .'' arXiv. https://doi.org/10.48550/arXiv.2103.00020

    2021 doi arXiv

  36. [44]

    Ramesh, Aditya, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. ``Zero- Shot Text -to- Image Generation .'' arXiv. https://doi.org/10.48550/arXiv.2102.12092

    2021 doi arXiv

  37. [45]

    Rombach, Robin, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. ``High- Resolution Image Synthesis with Latent Diffusion Models .'' arXiv. https://doi.org/10.48550/arXiv.2112.10752

    2022 doi arXiv

  38. [46]

    Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. 2015. ``U- Net : Convolutional Networks for Biomedical Image Segmentation .'' arXiv. https://doi.org/10.48550/arXiv.1505.04597

    2015 doi arXiv

  39. [47]

    Saad, Muhammad Muneeb, Ruairi O'Reilly, and Mubashir Husain Rehmani. 2023. ``A Survey on Training Challenges in Generative Adversarial Networks for Biomedical Image Analysis .'' arXiv. https://doi.org/10.48550/arXiv.2201.07646

    2023 doi arXiv

  40. [48]

    Sakai, Masato, Yiru Lai, Jorge Olano Canales, Masao Hayashi, and Kohhei Nomura. 2023. ``Accelerating the Discovery of New Nasca Geoglyphs Using Deep Learning.'' Journal of Archaeological Science 155 (July): 105777. https://doi.org/10.1016/j.jas.2023.105777

    2023

  41. [49]

    Simonyan, Karen, and Andrew Zisserman. 2015. ``Very Deep Convolutional Networks for Large - Scale Image Recognition .'' arXiv. https://doi.org/10.48550/arXiv.1409.1556

    2015 doi arXiv

  42. [50]

    Sinopoli, Carla M. 1991. Approaches to Archaeological Ceramics . Boston, MA: Springer US. https://doi.org/10.1007/978-1-4757-9274-4

    1991 doi

  43. [51]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Sohl-Dickstein, Jascha, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. ``Deep Unsupervised Learning Using Nonequilibrium Thermodynamics .'' arXiv. https://doi.org/10.48550/arXiv.1503.03585

    2015 doi arXiv

  44. [52]

    Steiner, Mélanie, and Lindsay Allason-Jones. 2005. Approaches to Archaeological Illustration: A Handbook. Practical Handbooks in Archaeology 18. York: Council for British Archaeology

    2005

  45. [53]

    Varghese, Meenu, Satheesh Raj, and Vigneshwaran Venkatesh. 2022. ``Influence of AI in Human Lives.'' arXiv. https://doi.org/10.48550/arXiv.2212.12305

    2022 doi arXiv

  46. [54]

    Vidale, M. 2007. Ceramica e Archeologia. 1. ed. Le Bussole Archeologia 285. Roma: Carocci

    2007

  47. [55]

    Wang, Yifei. 2023. ``Artificial Creativity - Ethical Reflections on AI 's Role in Artistic Endeavors .'' https://doi.org/10.36227/techrxiv.23897169

    2023 doi

  48. [56]

    Efros, Eli Shechtman, and Oliver Wang

    Zhang, Richard, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. 2018. ``The Unreasonable Effectiveness of Deep Features as a Perceptual Metric .'' arXiv. https://doi.org/10.48550/arXiv.1801.03924

    2018 doi arXiv

  49. [57]

    Zhang, Xinyi. 2024. `` AI - Assisted Restoration of Yangshao Painted Pottery Using LoRA and Stable Diffusion .'' Heritage 7 (11): 6282--6309. https://doi.org/10.3390/heritage7110295

    2024 doi

  50. [58]

    Zhou, Eric, and Dokyun Lee. 2024. ``Generative Artificial Intelligence, Human Creativity, and Art.'' PNAS Nexus 3 (3): pgae052. https://doi.org/10.1093/pnasnexus/pgae052. CSLReferences document

    2024 doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.