CoChat

Stable Diffusion Models Reveal a Persisting Human–AI Gap in Visual Creativity

Open access
Rondini, Silvia, Alvarez‐Martin, Claudia, Angermair‐Barkai, Paula, Penacchio, Olivier, Paz, Marc, Pelowski, Matthew, Dediu, Dan, Rodriguez‐Fornells, Antoni, Cerda‐Company, Xim · 2026 · Advanced Science

The DOI matches its Crossref record; 56 of the 56 references it cites were found in a registry or an index.

Abstract

While recent research suggests Large Language Models match human creative performance in divergent thinking tasks, visual creativity remains underexplored. This study compared image generation in human participants (Visual Artists and Non‐Artists) and using an image‐generation AI model (two prompting conditions with varying human input: high for Human‐Inspired, low for Self‐Guided). The creativity of the resulting images was evaluated by human raters (N = 255) and GPT‐4o acting as an AI rater under two conditions: strictly mirroring the human rating task and using in‐context learning with human‐rated examples as guidance. We observed a clear creativity gradient: Visual Artists > Non‐Artists ≥ Human‐Inspired GenAI > Self‐Guided GenAI. Increased human guidance strongly improved GenAI's creative output, bringing its productions close to those of Non‐Artists. Moreover, while Guided‐GPT‐4o more closely approximated human creativity judgment patterns, baseline GPT‐4o (without guidance) exhibited markedly different creativity evaluations, showing reduced discrimination between image categories and inflated scores for GenAI outputs. These results suggest that, in contrast to language‐centered tasks, GenAI models may face unique challenges in visual domains, where creativity depends on perceptual nuance and contextual sensitivity, distinctly human capacities that may not be readily transferable from language models.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1002/advs.202524142

Cites

56 of 56 references verified0 citation errors · 0 not checkable

Statistics

Figures

FIGURE 1 Scheme of the study's four phases. Phase I: Creative image generation by Visual‐Artists and Non‐Artists. The resulting human images were used to fine‐tune the Diffusion model (SDXL) through Low‐Rank Adaptation (LoRA), while the human‐generated ideas were used in the prompt of the Human‐Inspired GenAI group. Phase II: Creative image generation by GenAI (Human‐Inspired and Self‐Guided). Phase III: Creative image dataset rating by human raters. Phase IV: Creative image dataset rating by both GPT‐4o and Guided‐GPT‐4o (using in‐context learning).
FIGURE 1 Scheme of the study's four phases. Phase I: Creative image generation by Visual‐Artists and Non‐Artists. The resulting human images were used to fine‐tune the Diffusion model (SDXL) through Low‐Rank Adaptation (LoRA), while the human‐generated ideas were used in the prompt of the Human‐Inspired GenAI group. Phase II: Creative image generation by GenAI (Human‐Inspired and Self‐Guided). Phase III: Creative image dataset rating by human raters. Phase IV: Creative image dataset rating by both GPT‐4o and Guided‐GPT‐4o (using in‐context learning).CC BY · PubMed Central
FIGURE 2 Examples of drawings from each category (in rows): Visual Artists, Non‐Artists, Human‐Inspired GenAI, Self‐Guided GenAI. The values reported in the right‐hand top corner of each drawing correspond to the drawing's Creativity score by humans, GPT‐4o, and Guided‐GPT‐4o raters, respectively.
FIGURE 2 Examples of drawings from each category (in rows): Visual Artists, Non‐Artists, Human‐Inspired GenAI, Self‐Guided GenAI. The values reported in the right‐hand top corner of each drawing correspond to the drawing's Creativity score by humans, GPT‐4o, and Guided‐GPT‐4o raters, respectively.CC BY · PubMed Central
FIGURE 3 Creativity Ratings by Humans and GPT‐4o. A. Mean Creativity score (1–7) for each image category (Visual Artists, Non‐Artists, Human‐Inspired GenAI, and Self‐Guided)GenAI). B. Mean ratings for each Creativity dimension (Liking, Vividness, Originality, Aesthetics, Curiosity) by image category. C. Mean Creativity score for each image category based on GPT‐4o ratings. D. Mean ratings for each Creativity dimension by image category based on GPT‐4o evaluations. E. Mean Creativity score for each image category based on Guided‐GPT‐4o ratings. F. Mean ratings for each Creativity dimension by image category based on Guided‐GPT‐4o evaluations. Individual points represent image means, and error bars represent category mean ± standard deviation. Pairwise comparisons from linear mixed‐effect models are shown for adjacent categories. * p <.05, ** p <.01, *** p <.001. n.s. = not significant.
FIGURE 3 Creativity Ratings by Humans and GPT‐4o. A. Mean Creativity score (1–7) for each image category (Visual Artists, Non‐Artists, Human‐Inspired GenAI, and Self‐Guided)GenAI). B. Mean ratings for each Creativity dimension (Liking, Vividness, Originality, Aesthetics, Curiosity) by image category. C. Mean Creativity score for each image category based on GPT‐4o ratings. D. Mean ratings for each Creativity dimension by image category based on GPT‐4o evaluations. E. Mean Creativity score for each image category based on Guided‐GPT‐4o ratings. F. Mean ratings for each Creativity dimension by image category based on Guided‐GPT‐4o evaluations. Individual points represent image means, and error bars represent category mean ± standard deviation. Pairwise comparisons from linear mixed‐effect models are shown for adjacent categories. * p <.05, ** p <.01, *** p <.001. n.s. = not significant.CC BY · PubMed Central
FIGURE 4 Similarities and differences between Human and GPT‐4o ratings. A. Pearson correlations between Human and GPT‐4o across the Creativity score and the five rated subscales. B. Pearson correlations between Human and Guided‐GPT‐4o ratings across the Creativity score and the five rated subscales. C. Boxplot showing Human vs GPT‐4o overall Creativity ratings across image categories (Visual Artists, Non‐Artists, HI‐GenAI, and SG‐GenAI).
FIGURE 4 Similarities and differences between Human and GPT‐4o ratings. A. Pearson correlations between Human and GPT‐4o across the Creativity score and the five rated subscales. B. Pearson correlations between Human and Guided‐GPT‐4o ratings across the Creativity score and the five rated subscales. C. Boxplot showing Human vs GPT‐4o overall Creativity ratings across image categories (Visual Artists, Non‐Artists, HI‐GenAI, and SG‐GenAI).CC BY · PubMed Central
FIGURE 5 A. Distribution of the responses among the k‐means with k = 2 clusters. Left: by rater type (colour, Human vs. GPT‐4o vs. Guided‐GPT‐4o). Middle: by rater type (colour) and drawer type (Human or Stable Diffusion). Right: by rater type (colour) and image category (Non‐Artists, Visual Artists, Human‐Inspired GenAI, and Self‐Guided GenAI). B. The one‐factor model fitted to the whole dataset, split by rater type (Human vs. GPT‐4o and Guided‐GPT‐4o). There are three arrows and loadings for each indicator, the first being for the human raters, the second for the GPT‐4o rater, and the third for the Guided‐GPT‐4o rater. The loadings are standardized, and all of them were significant (all ps <.001). C. Predicted Creativity score from a beta regression on the interaction between rater type and image category while controlling (as random effects) for item and individual rater.
FIGURE 5 A. Distribution of the responses among the k‐means with k = 2 clusters. Left: by rater type (colour, Human vs. GPT‐4o vs. Guided‐GPT‐4o). Middle: by rater type (colour) and drawer type (Human or Stable Diffusion). Right: by rater type (colour) and image category (Non‐Artists, Visual Artists, Human‐Inspired GenAI, and Self‐Guided GenAI). B. The one‐factor model fitted to the whole dataset, split by rater type (Human vs. GPT‐4o and Guided‐GPT‐4o). There are three arrows and loadings for each indicator, the first being for the human raters, the second for the GPT‐4o rater, and the third for the Guided‐GPT‐4o rater. The loadings are standardized, and all of them were significant (all ps <.001). C. Predicted Creativity score from a beta regression on the interaction between rater type and image category while controlling (as random effects) for item and individual rater.CC BY · PubMed Central
FIGURE 6 Impact of AReA scores on Creativity ratings. High‐AReA participants assign overall higher creativity ratings than low‐AReA participants. The pattern of discrimination between image categories remains constant.
FIGURE 6 Impact of AReA scores on Creativity ratings. High‐AReA participants assign overall higher creativity ratings than low‐AReA participants. The pattern of discrimination between image categories remains constant.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 11

  1. F.Goes P.Sawicki M.Grzes M.Volpe andJ.Watson “Pushing GPT's Creativity to Its Limits: Alternative Uses and Torrance Tests ” inProceedings of the 14th International Conference for Computational Creativity(Association for Computational Creativity 2023) 342–346.
  2. 10.52202/068431-1522 · DOI
  3. V.Boutin T.Fel andL.Singhal “Diffusion Models as Artists: Are We Closing the Gap between Humans and Machines?” inProceedings of the 40th International Conference on Machine Learning(PMLR 2023) 2953–3002.
  4. Boutin V., 2024, Advances in Neural Information Processing Systems
  5. Sternberg R. J., 1999, Handbook of Creativity, 3
  6. Kearney R., 2005, The Wake of Imagination: Toward a Postmodern Culture
  7. Montell C., 2002, On Evolution of God‐Seeking Mind: An Inquiry into Why Natural Selection Would Favor Imagination and Distortion of Sensory Experience, Evolutionary Cognition, 8, 1
  8. Vyshedskiy A., 2019, Neuroscience of Imagination and Implications for Human Evolution, Current Neurobiology, 10, 89
  9. Mithen S., 1996, The Prehistory of the Mind: A Search for the Origins of Art, Religion and Science
  10. Pfeiffer J. E., 1982, The Creative Explosion: An Enquiry Into the Origins of Art and Religion
  11. Schuhmann C., 2022, Advances in Neural Information Processing Systems

Metadata

Authors
Rondini, S., Alvarez‐Martin, C., Angermair‐Barkai, P., Penacchio, O., Paz, M., Pelowski, M., Dediu, D., Rodriguez‐Fornells, A., & Cerda‐Company, X.
Affiliations
  • Institut d'Investigació Biomédica de Bellvitge (ES)
  • Cognition and Brain Plasticity Unit Bellvitge Biomedical Research Institute L'Hospitalet De Llobregat Spain
  • Vienna Cognitive Science Hub Vienna Austria
  • Universitat Autònoma de Barcelona (ES)
  • Computer Vision Center (ES)
  • University of Vienna (AT)
  • Institució Catalana de Recerca i Estudis Avançats (ES)
  • Bridging Research in AI and Neuroscience Computer Vision Center Bellaterra Spain
Venue
Advanced Science
Publisher
Wiley
Published
2026-03-24
Volume
13
Issue
27
Pages
e24142-e24142
ISSN
2198-3844 (ISSN-L)
OpenAlex
W4416666700
Cited by
2
References
43
Citations by year
2026: 2
FWCI
5.62
Type
journal article
Language
en
Open access
Licence
Locations
  • Advanced Science (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • HAL (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
  • Agència de Gestió d'Ajuts Universitaris i de Recerca (SDC007/25/000029, SDC007/25/000005)
  • H2020 Marie Skłodowska-Curie Actions (945408)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Creativity in Education and Neuroscience
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Aesthetic Perception and Analysis
Keywords
visual creativity, image generation, large language models, GPT-4o, creativity assessment, in-context learning, human guidance, divergent thinking
MeSH terms
Large Language Models, Generative Artificial Intelligence, Adult, Artificial Intelligence, Creativity, Female, Humans, Male, Young Adult

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.