Pith. sign in

REVIEW 5 major objections 6 minor 70 references

Speaking images. A novel framework for the automated self-description of artworks

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A chained pipeline of four open-source models can turn a digitized artwork into a short video in which its main character speaks about the work.

desk verdict A modest, honest proof-of-concept for auto-generating 'speaking image' videos of artworks; the core idea works on frontal-face portraits, but the evaluation is anecdotal and the 'foundational layer' conclusion outruns the evidence. read the letter →

arxiv 2506.05368 v1 pith:R7CF2WNO submitted 2025-05-28 cs.CV cs.AI

Valentine Bernasconi , Gustavo Marfia This is my paper
classification cs.CVcs.AI
keywords speakingimagegenerativeAIdigitalculturalheritageartworkdescriptionfaceanimationaudio-drivenportraitlargelanguagemodelsdeepfakes
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

This paper proposes that a digitized artwork can be made to explain itself: a short video in which a depicted character visibly talks about the work, produced automatically. The authors chain four open-source components—a multimodal language model, a face detector, a text-to-speech model, and an audio-driven portrait animator—and test the chain on fifteen artworks spanning paintings, religious scenes, and photographs. The central claim is that this concatenation already works well enough on frontal, detectable faces to serve as a 'primary and foundational layer' for engaging with digital cultural collections. If true, museum and heritage institutions could generate first-person narrated presentations of paintings at scale, while inheriting the biases and guardrail failures of the underlying models.

What carries the argument

The load-bearing object is the four-stage pipeline, which the paper calls the 'speaking image'—a video assembled from the original artwork plus a synthetic animated face and synchronized audio. Stage one runs a multimodal large language model (Llama 3.2) on a first-person prompt that asks for a two-sentence description 'as if the [female|male] character was speaking'; stage two detects faces and their gender with Deepface using the MTCNN backbone; stage three converts the text into gendered speech with Kokoro; stage four feeds the cropped, squared face and the audio into Hallo, which generates lip-synced animation; finally OpenCV reinserts the animated frames into the original image and MoviePy aligns audio. The pipeline is what carries the argument because every result, positive or negative, is attributed to the interaction of these stages.

What would settle it

Run the published pipeline on a corpus of artworks with controlled face orientations: if success concentrates in near-frontal faces and collapses beyond roughly 30 degrees of rotation, while non-figurative works produce no video at all, then the framework's operating envelope is exactly the face-animation bottleneck, not the language or speech stages.

Watch

Extended reading notes

Core claim

The paper's thesis is that the autonomous image is now technically achievable: given only a digitized artwork, the pipeline assembles a video in which the main character announces itself and describes the scene in the first person. On the paper's own terms, the discovery is that a straightforward concatenation of Llama 3.2 for image-to-text, Deepface with its MTCNN backbone for face detection and gender estimation, Kokoro for text-to-speech, and Hallo for audio-driven portrait animation produces coherent speaking images for many paintings, including non-realistic ones such as the naive self-portrait of Antonio Ligabue. The authors also report where the chain breaks: profile faces cannot be animated (the Christ figure in El Greco's 'The Baptism of Christ'), longer audio degrades output quality (median PSNR near 30 dB), and the language model either hallucinates attributions or refuses to speak when its guardrails trip, as with the photograph 'Alberto Burri, anni 70'. The intended contribution is the framework itself and the critical perspective on what the resulting 'speaking image' means for art history, education, and digitized heritage.

Load-bearing premise

The pipeline works only when the artwork contains a face that Deepface detects and that Hallo can animate, meaning a face near frontal with less than roughly 30 degrees of rotation; profile portraits, group scenes, small faces, and non-figurative works stall the chain or leave most of the content unspoken.

Editorial extensions

If this is right

  • Museum and heritage platforms could automatically generate short first-person video introductions for any digitized artwork whose main character has a detectable, near-frontal face.
  • Prompt design becomes an editorial lever: giving the model author, title, and date shifts results toward context and emotion for religious works, but triggers content-moderation refusals for photographs and unknown attributions.
  • The measured quality drop (median PSNR around 30 dB, FID of 293.67 against the GRID corpus) implies that long single-shot animations will need chunking or smoothing before deployment.
  • The framework positions the speaking image as a 'foundational layer' for augmented-reality encounters with under-documented frescoes and for assisted annotation of large digitized collections, but only if art historians curate the LLM output.
  • Both Llama-biased visual hallucination and Llama Guard 3 Vision's refusal behavior are framed as features the next iteration must handle through post-curation and prompt refinement.

Reading between the lines

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

  • Beyond the paper: the pipeline could be repurposed as a bias probe, since asking the same model to speak as characters of different genders or from different artistic traditions would expose asymmetries in its art-historical knowledge.
  • Beyond the paper: a fallback mode for face-less or group scenes—using the same LLM text and TTS audio with a caption overlay or off-screen voice—would extend the educational benefit to the large share of artworks the current chain cannot animate.
  • Beyond the paper: each completed speaking-image video is arguably a new version of the artwork rather than a transparent rendering of it, so museum use would need to decide how to label the artifact's provenance.
  • The authors leave implicit that the same four-model chain, with its failure cases, is effectively a measurement of how much of a painting's content survives a frontal-face bottleneck; quantifying that 'survival rate' over a larger corpus would be a direct next experiment.
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

5 major / 6 minor

Summary. The paper proposes and implements an automated pipeline that turns a digitized artwork into a short video in which a depicted character appears to speak a first-person description of the work. Four off-the-shelf components are chained: Deepface (with MTCNN backbone) for face detection and gender estimation, Llama 3.2 (via Ollama) for image-to-text narration, Kokoro for text-to-speech, and Hallo for audio-driven portrait animation; the animated face is then reinserted into the original image with OpenCV and synchronized with MoviePy. The authors test the pipeline on 15 artworks spanning early modern paintings, contemporary photography, a frame-back image, and a hand detail. They report qualitative analyses of LLM behavior (hallucinated attributions, guardrail refusals, emotional interpretations), a comparison of face detectors on 9 images, and quantitative quality metrics (PSNR, FID, FVD) for the animation stage. The paper discusses cultural biases, art-historical implications, and potential applications in museums and AR, and concludes that the "speaking image" is a foundational layer for engaging with digitized cultural heritage.

Significance. The strength of the paper is that it is a concrete, reproducible proof-of-concept: the code is publicly released, the pipeline is composed entirely of open-source components, and the authors document failure cases (profile faces, non-figurative images, guardrail refusals) honestly. The discussion of LLM bias and the personification of the model raises interesting questions for digital art history. However, the evaluation is anecdotal: the 15-image test set is small and curated, there is no user study or quantitative comparison to alternative approaches, and the reported FID/FVD scores are compared to published values from a different reference set. As a result, the central claim—that the framework automatically produces speech-animated artworks for a broad class of digitized artifacts—is not yet supported. If the authors add a larger-scale coverage analysis and temper the concluding claims, the work would be a useful contribution to the digital cultural heritage community.

major comments (5)
  1. [Section 3.3, Section 2.5, Section 4] The central claim that the pipeline automatically produces a speaking image for digitized artworks is only demonstrated on a curated set of 15 images, and the paper's own results show that the pipeline fails for profile faces (El Greco's 'The Baptism of Christ'), for non-figurative inputs (the back of a frame and a hand detail), and for at least one photograph to which Llama Guard 3 Vision refuses a response (Mario Dondero's portrait). Since the conclusion in Section 4 calls the framework 'a primary and foundational layer' for digitized cultural heritage, the authors need to either restrict this claim to the subclass of artworks with a detectable, near-frontal face, or provide evidence on the fraction of real museum collections that satisfy this condition. A larger-scale systematic evaluation with per-step success rates would be needed to support the current conclusion.
  2. [Section 3.2, Figure 2] The comparison of face detection models is based on only 9 images, and the paper does not state how these images were selected nor does it report the raw counts behind the confusion matrices. Without a defined protocol and quantitative precision/recall values, the claim that Deepface with the MTCNN backbone 'yields best results' is not reproducible. This is important because face detection is the first and most brittle step of the pipeline.
  3. [Section 3.3, Figure 4] The PSNR values computed between the original face crop and the last animation frame are reported with a median of about 30 dB, but no baseline or threshold is provided to interpret whether this reflects acceptable perceptual quality. To support the statement that 'we were able to produce good quality animations,' the authors should compare against another talking-head method on the same inputs or include a perceptual evaluation.
  4. [Section 3.3, FID/FVD paragraph] The FID and FVD scores are computed using the GRID corpus as a real-video reference and then compared to the scores reported in the Hallo paper [28]. Since the reference datasets and pipeline inputs are different, the comparison is not valid evidence of the animation quality relative to Hallo. The paper acknowledges the mismatch but nevertheless concludes that the quality 'seems consistent with the results promised by Hallo'; this inference does not follow. The authors should either run the Hallo baseline on the same reference set or explicitly state that no direct comparison is being made.
  5. [Section 2.1, Section 2.4] The pipeline does not specify how the 'main character' is chosen when multiple faces are detected. The LLM receives only the gender of the detected face and generates a text that may describe a figure other than the one animated; for example, in Caravaggio's 'The incredulity of Saint Thomas,' several male figures are present, and the speech could be attributed to the wrong character. The framework should either select the main character deterministically (e.g., largest or most central face) or generate and attach separate audio tracks to each detected face.
minor comments (6)
  1. [Section 2.2] The face detector is referred to as 'MtCnn' in the text, but the standard abbreviation is 'MTCNN'; please use consistent capitalization.
  2. [Section 3.3] The GRID corpus reference [10] is given as a Zenodo DOI only; please provide a full bibliographic entry (authors, year, venue).
  3. [Figure 2] The confusion-matrix caption is vague; please specify what is being compared (e.g., detected faces vs. manually annotated ground truth) and how true/false positives are defined for artworks.
  4. [Section 3.1.1] The quoted example from 'Eating Figures (Quick Snack)' is informal and not clearly marked as a verbatim model output; please format all LLM quotes consistently and indicate whether they are exact outputs or paraphrases.
  5. [Section 1, Section 3.1.4] The concept of the 'autonomous image' is central to the framing but is not defined; a brief definition or a pointer to the relevant literature would help readers.
  6. [Section 4] The paper would benefit from an explicit ethics statement given the extended discussion of deepfakes, privacy, and cultural bias.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline's outputs are evaluated against external benchmarks, and its documented failure modes are empirical coverage limits rather than reasoning that reduces to its inputs.

full rationale

The paper proposes an assembly pipeline (Llama + Deepface + Kokoro + Hallo) and evaluates it with PSNR, FID, and FVD. None of these metrics is fitted from the test set and then renamed as a prediction; the paper explicitly reports that its FID (293.67) is substantially worse than Hallo's published 44.578, which is the opposite of a self-serving equivalence. The central capability claim depends on Hallo's requirement of a near-frontal face (Section 3.3, El Greco's Christ in profile), but that is a robustness/coverage limitation, not a circular step: the failure is discovered by running the pipeline, not assumed by it. The authors' self-citation [4] concerns hand-pose recognition and is mentioned only as a future combination, so it is not load-bearing. No equation, fitted parameter, or uniqueness theorem is invoked to force the paper's conclusions. The 'speaking image' label is a descriptive term for the composite video output, not a re-derivation of a known result. Therefore no circularity is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The framework pulls almost all of its working parts from prior models. It introduces no new mathematical derivations and no fitted parameters. The main axioms are domain assumptions that the chosen models will transfer acceptably from photographs to paintings, and that a 15-image hand-picked set can support qualitative conclusions. The 'speaking image' is a framing concept, not an entity with independent evidence.

free parameters (1)
  • audio chunk duration limit = about 20 seconds
    Chosen by hand to limit Hallo output deformation; not derived from data. Mentioned in Section 3.1 as a constraint for the LLM prompt.
assumptions (4)
  • domain assumption Face detection and animation models perform adequately on painted faces
    The pipeline depends on Deepface/MTCNN and Hallo succeeding on artistic images; the paper tests on 9 and 15 images, and reports failures (e.g., profile face of Christ in El Greco, Section 3.3).
  • domain assumption Llama 3.2 can produce first-person descriptions in the requested voice and style
    The central output text comes from Llama with a two-sentence prompt; the paper reports hallucinations, refusals, and guardrail triggers (Sections 3.1.1-3.1.4).
  • domain assumption The 15-image sample is representative enough for the qualitative conclusions
    The test set is small and selected by the authors (Section 2.5); not a random sample.
  • standard math Standard image cropping and video assembly operations are reliable
    The bounding-box squaring and frame insertion use basic geometry and OpenCV operations; no special axioms are required.
invented entities (1)
  • Speaking image (animated artwork video)
    purpose: The output artifact of the pipeline: an artwork whose character animates and narrates its own content.
    The term is coined in this paper as a framing concept; it is not a scientific entity with independent falsifiable predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speaking images. A novel framework for the automated self-description of artworks." pith.science (2026). https://pith.science/paper/R7CF2WNO

@misc{pith2026250605368,
  author       = {Pith},
  title        = {Pith review of: Speaking images. A novel framework for the automated self-description of artworks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7CF2WNO}},
  note         = {Machine review of arXiv:2506.05368}
}
read the original abstract

Recent breakthroughs in generative AI have opened the door to new research perspectives in the domain of art and cultural heritage, where a large number of artifacts have been digitized. There is a need for innovation to ease the access and highlight the content of digital collections. Such innovations develop into creative explorations of the digital image in relation to its malleability and contemporary interpretation, in confrontation to the original historical object. Based on the concept of the autonomous image, we propose a new framework towards the production of self-explaining cultural artifacts using open-source large-language, face detection, text-to-speech and audio-to-animation models. The goal is to start from a digitized artwork and to automatically assemble a short video of the latter where the main character animates to explain its content. The whole process questions cultural biases encapsulated in large-language models, the potential of digital images and deepfakes of artworks for educational purposes, along with concerns of the field of art history regarding such creative diversions.

Figures

Figures reproduced from arXiv: 2506.05368 by the authors.

Figure 1
Figure 1. Pipeline for the automated creation of a speaking image. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Confusion matrix comparison for face detection models. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Two examples of animated faces where deformations appear over time. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Boxplot chart representing the distribution of PSNR values. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 34 canonical work pages

  1. [28]

    Mingwang Xu et al. 2024. Hallo: Hierarchical Audio-Driven Visual Synthesis for Portrait Image Animation. arXiv:2406.08801 [cs.CV] https: //arxiv.org/abs/2406.08801

    arXiv 2024

  2. [1]

    Badr AlKhamissi and et al. 2024. Investigating Cultural Alignment of Large Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Bangkok, Thailand, 12404–12422. arXiv:2402.13231 [cs.CL] https://arxiv.org/abs/2402.13231

    arXiv 2024

  3. [2]

    David Baidoo-Anu and Leticia Owusu Ansah. 2023. Education in the era of generative artificial intelligence (AI): Understanding the potential benefits of ChatGPT in promoting teaching and learning.Journal of AI7, 1 (2023), 52–62

    2023

  4. [3]

    2017.Beyond the Frame: Case Studies

    Dominique Bauer. 2017.Beyond the Frame: Case Studies. ASP Editions, Escala, France. https://www.libreriauniversitaria.it/beyond-frame-case- studies-bauer/book/9789057184833

    work page arXiv 2017

  5. [4]

    Valentine Bernasconi, Eva Cetinić, and Leonardo Impett. 2023-06. A Computational Approach to Hand Pose Recognition in Early Modern Paintings. Journal of Imaging9, 6 (2023-06), 120. doi:10.3390/jimaging9060120

    work page doi:10.3390/jimaging9060120 2023

  6. [5]

    Federico et al. Bianchi. 2023. Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. InProceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’23). Association for Computing Machinery, New York, NY, USA, 1493–1504. doi:10.1145/3593013.3594095

    arXiv 2023

  7. [6]

    Yi et al. Bin. 2024. GalleryGPT: Analyzing Paintings with Large Multimodal Models. InProceedings of the 32nd ACM International Conference on Multimedia (MM ’24). ACM, New York, USA, 7734–7743. doi:10.1145/3664647.3681656

    arXiv 2024

  8. [7]

    2022.Lire l’œuvre d’Aby Warburg à la lumière de ses Fragments sur l’expression

    Lara Bonneau. 2022.Lire l’œuvre d’Aby Warburg à la lumière de ses Fragments sur l’expression. Les presses du réel, Dijon, France. https: //www.lespressesdureel.com/ouvrage.php?id=8997&menu=0

    work page 2022

Show all 70 references
  1. [8]

    Chakrabarty

    Tuhin et al. Chakrabarty. 2024. Art or Artifice? Large Language Models and the False Promise of Creativity. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 30, 34 pages. d...

    2024

  2. [9]

    Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. DALL-EVAL: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models . In2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, Los Alamitos, CA, USA, 3020–3031. doi:10.11...

    2023

  3. [10]

    2006.The Grid Audio-Visual Speech Corpus

    Martin Cooke and et al. 2006.The Grid Audio-Visual Speech Corpus. doi:10.5281/zenodo.3625687

    2006 doi

  4. [11]

    Fatemeh Daneshfar, Ako Bartani, and Pardis Lotfi. 2024. Image captioning by diffusion models: A survey.Engineering Applications of Artificial Intelligence138 (2024), 109288. doi:10.1016/j.engappai.2024.109288

    2024

  5. [12]

    2023-06-01

    Nassim Dehouche and Kullathida Dehouche. 2023-06-01. What’s in a text-to-image prompt? The potential of stable diffusion in visual arts education. Heliyon9, 6 (2023-06-01). doi:10.1016/j.heliyon.2023.e16757 Publisher: Elsevier

    2023 doi

  6. [13]

    Georges Didi-Huberman. 2003. Before The Image, Before Time: The Sovereignty of Anachronism. InCompelling Visuality: The Work Of Art In And Out Of History(first edition ed.), Claire Farago and Robert Zwijnenberg (Eds.). University Of Minnesota Press, Minneapolis, USA

    2003

  7. [14]

    Temenuzhka Dimova. 2023. Chiroscript: Transcription System for Studying Hand Gestures in Early Modern Painting.Arts12, 4 (2023), 179. doi:10.3390/arts12040179

    2023 doi

  8. [15]

    1999.Painting the word: Christian pictures and their meanings

    John Drury. 1999.Painting the word: Christian pictures and their meanings. Yale University Press, New Haven, USA. Pages: 201 Repository: aleph.mpg.de Library Catalog

    1999

  9. [16]

    So what if ChatGPT wrote it?

    Yogesh K Dwivedi and et al. 2023. Opinion Paper:“So what if ChatGPT wrote it?” Multidisciplinary perspectives on opportunities, challenges and implications of generative conversational AI for research, practice and policy.International journal of information management71 (2023...

    2023

  10. [17]

    Bengamra S. et al. 2024. A comprehensive survey on object detection in Visual Art: taxonomy and challenge.Multimed Tools Application83 (2024), 14637–14670. doi:10.1007/s11042-023-15968-9

    2024 doi

  11. [18]

    Cheng Li et al. 2024. CulturePark: Boosting Cross-cultural Understanding in Large Language Models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems. arXiv:2405.15145 [cs.AI] https://arxiv.org/abs/2405.15145

    2024 arXiv

  12. [19]

    Danny Driess et al. 2023. PaLM-E: An Embodied Multimodal Language Model. arXiv:2303.03378 [cs.LG] https://arxiv.org/abs/2303.03378

    2023 arXiv

  13. [20]

    Edresson Casanova et al. 2024. XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model. arXiv:2406.04904 [eess.AS] https://arxiv.org/abs/ 2406.04904

    2024 arXiv

  14. [21]

    Hannah Lee et al. 2024. The Tug-of-War Between Deepfake Generation and Detection. arXiv:2407.06174 [cs.CV] https://arxiv.org/abs/2407.06174

    2024 arXiv

  15. [22]

    Jianfeng Chi et al. 2024. Llama Guard 3 Vision: Safeguarding Human-AI Image Understanding Conversations. arXiv:2411.10414 [cs.CV] https: //arxiv.org/abs/2411.10414

    2024 arXiv

  16. [23]

    Jianzhu Guo et al. 2025. LivePortrait: Efficient Portrait Animation with Stitching and Retargeting Control. arXiv:2407.03168 [cs.CV] https: //arxiv.org/abs/2407.03168

    2025 arXiv

  17. [24]

    Jianfeng Wang et al. 2022. GIT: A Generative Image-to-text Transformer for Vision and Language. arXiv:2205.14100 [cs.CV] https://arxiv.org/abs/ 2205.14100

    2022 arXiv

  18. [25]

    Jiannan Xiang et al. 2023. Language Models Meet World Models: Embodied Experiences Enhance Language Models. arXiv:2305.10626 [cs.CL] https://arxiv.org/abs/2305.10626

    2023 arXiv

  19. [26]

    Kazuki Hayashi et al. 2024. Artwork Explanation in Large-scale Vision Language Models. arXiv:2403.00068 [cs.CV] https://arxiv.org/abs/2403.00068

    2024

  20. [27]

    Mohammadreza Farrokhnia et al. 2024. A SWOT analysis of ChatGPT: Implications for educational practice and research.Innovations in Education and Teaching International61, 3 (2024), 460–474. doi:10.1080/14703297.2023.2195846 arXiv:https://doi.org/10.1080/14703297.2023.2195846

    2024

  21. [29]

    Qiujie Xie et al. 2025. Human Simulacra: Benchmarking the Personification of Large Language Models. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=BCP5nAHXqs

    2025

  22. [30]

    Rui Sun et al. 2024. From Sora What We Can See: A Survey of Text-to-Video Generation. arXiv:2405.10674 [cs.CV] https://arxiv.org/abs/2405.10674

    2024 arXiv

  23. [31]

    Sanyuan Chen et al. 2024. VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers. arXiv:2406.05370 [cs.CL] https://arxiv.org/abs/2406.05370

    2024 arXiv

  24. [32]

    Shuo Wang et al. 2025. Exploring the Impact of Personality Traits on LLM Bias and Toxicity. arXiv:2502.12566 [cs.AI] https://arxiv.org/abs/2502.12566

    2025

  25. [33]

    Xinsheng Wang et al. 2025. Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens. arXiv:2503.01710 [cs.SD] https://arxiv.org/abs/2503.01710

    2025 arXiv

  26. [34]

    Yi Dong et al. 2024. Building Guardrails for Large Language Models. arXiv:2402.01822 [cs.CL] https://arxiv.org/abs/2402.01822

    2024 arXiv

  27. [35]

    Yi Dong et al. 2024. Safeguarding Large Language Models: A Survey. arXiv:2406.02622 [cs.CR] https://arxiv.org/abs/2406.02622

    2024 arXiv

  28. [36]

    Francis, Sue Jones, and David P

    Nigel J. Francis, Sue Jones, and David P. Smith. 2025. Generative AI in Higher Education: Balancing Innovation and Integrity.British Journal of Biomedical ScienceVolume 81 - 2024 (2025). doi:10.3389/bjbs.2024.14048

    2025

  29. [37]

    Aaron Grattafiori and al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783 [cs.AI] https://arxiv.org/abs/2407.21783

    2024 arXiv

  30. [38]

    Yiwei Guo and et al. 2024. VoiceFlow: Efficient Text-To-Speech with Rectified Flow Matching. InICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 11121–11125. doi:10.1109/ICASSP48485.2024.10445948

    2024

  31. [39]

    Serengil Sefik Ilkin and Ozpinar Alper. 2024. A Benchmark of Facial Recognition Pipelines and Co-Usability Performances of Modules.Bilisim Teknolojileri Dergisi17, 2 (2024), 95–107. doi:10.17671/gazibtd.1399077

    2024 doi

  32. [40]

    Harry H. et al. Jiang. 2023. AI Art and its Impact on Artists. InProceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’23). Association for Computing Machinery, New York, NY, USA, 363–374. doi:10.1145/3600211.3604681 Speaking images. A novel framework f...

    2023

  33. [41]

    2023-04-05

    Xuan Ju and et al. 2023-04-05. Human-Art: A Versatile Human-Centric Dataset Bridging Natural and Artificial Scenes. doi:10.48550/arXiv.2303.02760 arXiv:2303.02760

    2023 doi arXiv

  34. [42]

    Steffen Jung and Margret Keuper. 2021. Internalized Biases in Fréchet Inception Distance. InNeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications. https://openreview.net/forum?id=mLG96UpmbYz

    2021

  35. [43]

    2023.Animation between magic, miracles and mechanics : principles of life in medieval imagery

    Hans Henrik Lohfert Jørgensen. 2023.Animation between magic, miracles and mechanics : principles of life in medieval imagery. Aarhus University Press, Aarhus, Denmark

    2023

  36. [44]

    Amina Kammoun and et al. 2022. Generative Adversarial Networks for Face Generation: A Survey.ACM Comput. Surv.55, 5, Article 94 (2022), 37 pages. doi:10.1145/3527850

    2022 doi

  37. [45]

    Afshin Khadangi and et al. 2025. CognArtive: Large Language Models for Automating Art Analysis and Decoding Aesthetic Elements. arXiv:2502.04353 [cs.CL] https://arxiv.org/abs/2502.04353

    2025 arXiv

  38. [46]

    2024.How I smoothly injected a Mona Lisa talking-head into the full portrait

    Mohammed El komy. 2024.How I smoothly injected a Mona Lisa talking-head into the full portrait. https://medium.com/analytics-vidhya/how-i- smoothly-injected-a-mona-lisa-talking-head-into-the-full-portrait-60f3f5e4b74b

    2024

  39. [47]

    Yinghao Aaron Li, Cong Han, and Nima Mesgarani. 2025. StyleTTS: A Style-Based Generative Model for Natural and Diverse Text-to-Speech Synthesis.IEEE Journal of Selected Topics in Signal Processing19, 1 (2025), 283–296. doi:10.1109/JSTSP.2025.3530171

    2025

  40. [48]

    Mingxiang et al. Liao. 2024. Evaluation of Text-to-Video Generation Models: A Dynamics Perspective. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37. Curran Associates, Inc., ...

    2024

  41. [49]

    Mohod Madhu. 2025. Ai as Artist: The Role of Generative Algorithms in Creative Industries.International Journal of Advanced Research in Education and Technology12, 2 (2025), 542–548

    2025

  42. [50]

    2021-08-01

    Mihaela Mihailova. 2021-08-01. To Dally with Dalí: Deepfake (Inter)faces in the Art Museum. 27, 4 (2021-08-01), 882–898. doi:10.1177/ 13548565211029401 Publisher: SAGE Publications Ltd

    2021

  43. [51]

    2024-05-27

    Mohammad Nadeem and et al. 2024-05-27. Vision-Enabled Large Language and Deep Learning Models for Image-Based Emotion Recognition | Cognitive Computation.Cognitive Computation16 (2024-05-27), 2566–2579. doi:10.1007/s12559-024-10281-5

    2024 doi

  44. [52]

    2013.Visual Time: The Image in History

    Keith Moxey. 2013.Visual Time: The Image in History. Duke University Press. doi:10.2307/j.ctv11smn17

    2013 doi

  45. [53]

    Lynda Nead. 1990. The Female Nude: Pornography, Art, and Sexuality.Signs15, 2 (1990), 323–335. https://www.jstor.org/stable/3174488

    1990

  46. [54]

    Russell S. A. Palmer and Andrew M. A. Palmer. 2025. Intention Is All You Need to Create Your Own Hollywood Blockbuster Movies—Revolutionizing Filmmaking With Generative AI and Saga.Computer58, 4 (2025), 79–94. doi:10.1109/MC.2025.3534263

    2025

  47. [55]

    Maria-Teresa De Rosa Palmini and Eva Cetinic. 2024. Patterns of Creativity: How User Input Shapes AI-Generated Visual Diversity. arXiv:2410.06768 [cs.HC] https://arxiv.org/abs/2410.06768

    2024 arXiv

  48. [56]

    Vinodkumar Prabhakaran, Rida Qadri, and Ben Hutchinson. 2022. Cultural Incongruencies in Artificial Intelligence. arXiv:2211.13069 [cs.CY] https://arxiv.org/abs/2211.13069

    2022 arXiv

  49. [57]

    Wenke Qin and et al. 2023. NeoDescriber: An image-to-text model for automatic style description of neoclassical architecture.Expert Systems with Applications231 (2023), 120706. doi:10.1016/j.eswa.2023.120706

    2023

  50. [58]

    Mansi Rehaan, Nirmal Kaur, and Staffy Kingra. 2024. Face manipulated deepfake generation and recognition approaches: a survey.Smart Science12, 1 (2024), 53–73. doi:10.1080/23080477.2023.2268380 arXiv:https://doi.org/10.1080/23080477.2023.2268380

    2024

  51. [59]

    Alberto Sanchez-Acedo and et al. 2024. The influence of audiovisual elements on the realism of generative AI videos: the case of Sora.Proceedings of the International Workshop on Artificial Intelligence and Creativity (CREAI), co-located with ECAI 2024(2024). https://cris.unib...

    2024

  52. [60]

    Robin Sheets. 1988. Pornography and Art: The Case of "Jenny".Critical Inquiry14, 2 (1988), 315–334. https://www.jstor.org/stable/1343449 Publisher: The University of Chicago Press

    1988

  53. [61]

    2022.Painting, History and Meaning: Sites of Time

    Craig Staff. 2022.Painting, History and Meaning: Sites of Time. Intellect Ltd. https://press.uchicago.edu/ucp/books/book/distributed/P/bo68883774. html

    2022

  54. [62]

    C. A. Strine, Mark Mcinroy, and Alexis Torrance (Eds.). 2022-08-11.Image As Theology: The Power of Art in Shaping Christian Thought, Devotion, and Imagination. Brepols Pub, Turnhout, Belgium

    2022

  55. [63]

    Xiaoguang et al. Tu. 2022. Image-to-Video Generation via 3D Facial Dynamics.IEEE Transactions on Circuits and Systems for Video Technology32, 4 (2022), 1805–1819. doi:10.1109/TCSVT.2021.3083257

    2022

  56. [64]

    Valencia

    Jackeline et al. Valencia. 2024-04-15. Using machine learning to predict artistic styles: an analysis of trends and the research agenda.Artificial Intelligence Review57, 5 (2024-04-15), 118. doi:10.1007/s10462-024-10727-0

    2024 doi

  57. [65]

    Vilchis and et al

    C. Vilchis and et al. 2023. A survey on the pipeline evolution of facial capture and tracking for digital humans.Multimedia Systems29 (2023), 1917–1940. doi:10.1007/s00530-023-01081-2

    2023 doi

  58. [66]

    Cong et al. Wang. 2025. DreamVideo: High-Fidelity Image-to-Video Generation with Image Retention and Text Guidance. InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 1–5. doi:10.1109/ICASSP49660.2025.10887583

    2025

  59. [67]

    Huawei Wei, Zejun Yang, and Zhisheng Wang. 2024. AniPortrait: Audio-Driven Synthesis of Photorealistic Portrait Animation. arXiv:2403.17694 [cs.CV] https://arxiv.org/abs/2403.17694

    2024 arXiv

  60. [68]

    Li L Wu JX. 2025. AI-driven computational creativity in fashion design: a review.Textile Research Journal95 (2025). doi:10.1177/00405175241279976

    2025 doi

  61. [69]

    Jiarui et al. Yu. 2023. CgT-GAN: CLIP-guided Text GAN for Image Captioning. InProceedings of the 31st ACM International Conference on Multimedia (MM ’23). Association for Computing Machinery, New York, NY, USA, 2252–2263. doi:10.1145/3581783.3611891 16 Bernasconi et al

    2023

  62. [70]

    Wenxuan et al. Zhang. 2023. SadTalker: Learning Realistic 3D Motion Coefficients for Stylized Audio-Driven Single Image Talking Face Animation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8652–8661

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.