1 Introduction↩︎

Over the past decade, Computational Creativity (CC) has experienced rapid growth, largely driven by advances in AI, in particular transformer models, which greatly increased the capacity of computer systems to create original artifacts.

In this work, we focus on a type of visual computer creativity inherently combinatorial in nature, defined in [1] as the process of combining ideas from a given conceptual space in novel ways, which minds both convergent creativity (the capacity to identify the best or correct solution to a problem) and divergent creativity (the capacity to generate a wide range of varied original responses) [2], in an assisted creativity framework, defined in [3] as the automatic process of supporting a human creator. In particular, we asked the Gemini Pro 3 Nano Banana model [4] to generate pastiches after contemporary art, via prompting, by uploading the images of the original artworks.

The starting point was the work of [5], who employed Chat GPT to create pastiches after the artworks of 12 contemporary artists. The study uncovered that the multi-faceted concept of style can be expressed by features like texture and color, semantics and concept, artistic touches, visual composition, or perceptions, and that these features can be captured independently using different computer vision models which extract high-dimensional embeddings. A large cosine distance between the AI generated pastiche and the original human artwork in these embedding spaces means that the pastiche was original, but satisfies less the task requirement of creating a pastiche, while a low distance indicates a copy that is less original.

We kept the same framework and data, but we used the newer SOTA image generation model Gemini. We were motivated by the following research questions:

1. Are newer generative models getting “better” at producing pastiches after human artworks?

2. How do human and automatic evaluation correlate in this respect?

3. Are visual models consistent in extracting stylistic features across different generative AI models?

Our study directly compares the outputs of the two models and gains insight into the stability of the claims made in [5]. A vital part of our assessment was the contribution of the artists themselves through their evaluation and comparison.

2 Related Work↩︎

Recently, the fast growing body of work in CC spread to include a wide variety of topics, from improvement techniques and evaluation to ethical concerns and the role of humans in automatic image generation.

For example, [6] showed that the use of more advanced prompt engineering techniques such as Optimization as PROmpting (OPRO) or chain-of-thought (COT) leads to no significant improvements in textual computational creative systems, while [7] analyzed how ChatGPT amplifies positive sentiment and adds descriptive language regardless of the user’s prompt style, showing no correlation between user prompt and AI-generated creative output.

The creative abilities of three artificial image-to-image generative models are formally quantified in [8]. The authors propose three objective creativity criteria: requirement satisfaction, cohesion, and novelty and introduce formal measures for each of them. The complexity and subjectivity of the evaluation process of CC systems, which depends mainly on the expertise and perception of the evaluator, are recognized in [1], [9], [10]. Moreover, [6], [11] point out that the domain is extremely diverse, making objective human evaluation difficult, time consuming, and costly.

Finally, [12] documented how artists negotiate authorship and ownership when experimenting with artificially generated imagery, revealing how they re-frame artistic roles and authorship through generative tools.

3 Data↩︎

This study uses the dataset proposed in [5], which consists of 108 images:

• 36 original artworks by twelve contemporary artists, who generously agreed to lend 3 of their artworks each, for research purposes. The artists are: Adi Matei, Ciprian Mureșan, Ion Grigorescu, Iulia Uță, Karine Fauchard, Lazar Lyutakov, Marius Tănăsescu, Mathias Poeschl, Oana Năstăsache, Philip Patkowitsch, Răzvan Botiș, and Tom Chamberlain;

• 72 AI pastiches produced by AI (ChatGPT 5 model), 2 imitations/pastiches for each original.

We enriched this image dataset with a new series of AI generated works (e.g., Figures 3, 4, 5, 6 from the Appendix) using the latest Gemini Pro 3 Nano Banana model [4]. We used the same prompt as in the ChatGPT experiments conducted by [5], so the results would be directly comparable.

4 Experimental Setup↩︎

In the pre-reprocessing step, we converted all images to RGB format and normalized them according to each of the five models used in [5] to extract high-dimensional embeddings and which have been shown to capture different aspects of artistic style, as follows:

• AdaIN-Style [13] captures 1920-dimensions of texture/color statistics;

• ResNet50-Style [14] captures 2048-dimensions of artistic style features;

• CLIP-ViT-L [15] captures 768-dimensions of semantic features;

• DINOv2 [16] captures 1024-dimensions of fine-grained visual features;

• VGG19 [17] captures 4096-dimensions of perceptual features.

To approximate the stylistic similarity corresponding to each of the five models, for each artwork group (the original and its two Gemini-generated pastiches), we compute three pairwise cosine distances: original to pastiche 1, original to pastiche 2, and pastiche 1 to pastiche 2.

5 Automatic evaluation↩︎

The distribution of the cosine distance between the original artworks and their Gemini-generated pastiches using the five distinct feature embedding models is given in Table 1, in comparison to the GPT results reported in [5]. Figure 1 illustrates these distances the Gemini generated pastiches, for a more intuitive visual representation.

Comparing ChatGPT and Gemini reveals a consistent model distance hierarchy: AdaIN-Style remains the lowest (0.064), followed by CLIP-ViT-L (0.167), ResNet50-Style (0.441), DINOv2 (0.461), and VGG19 (0.699). This confirms that the “compositional gap” between texture matching and perceptual similarity is model-agnostic and inherent to the task itself. However, the models prioritize different elements (Table 1). Gemini favors deep semantics, better capturing concepts (15.2% closer) and style (2.9%), making it more adept at the overall pasticheing task based on cosine distance. In contrast, ChatGPT focuses on surface forms, slightly better imitating perceptual features (3.7% closer) and texture/color (1.6%). Furthermore, Gemini generates significantly more diverse pastiches, as shown in Figure 10 from the Appendix. The average distance between pastiches of the same original is 10.1% higher for Gemini than GPT (0.374 vs. 0.340). This diversity is particularly notable in DINOv2 (+22.4%) and VGG19 (+10.3%), indicating that Gemini explores broader compositional and perceptual interpretations.

We also compared the discrimination (variance) and consistency of the two models, given in Table 2. Discrimination reflects the embedding model’s capacity to distinctly separate different artistic styles (measured by variance), while consistency indicates the stability of the generative model when producing multiple pastiches for the same original artwork (measured by the average distance between them).

DINOv2 remains the most discriminative model for both Gemini (0.0425) and GPT (0.0349), while AdaIN-Style remains the least discriminative. Notably, Gemini shows improved consistency on VGG19 and ResNet50-Style compared to GPT, suggesting more stable perceptual and stylistic reproduction.

The models capture complementary information, as evidenced by the weak correlation (\(r = 0.293\)) between DINOv2 (composition) and AdaIN-Style (texture). This quantifies the “compositional gap”, showing that texture fidelity does not predict compositional accuracy. Table 3 highlights some extreme cases in which these metrics disagree. For example, Ciprian Mureșan’s Work 1 shows the largest positive divergence (+0.798), pairing high texture adherence (AdaIN: 20th percentile) with high compositional divergence (DINOv2: 100th percentile). Conversely, Tom Chamberlain’s Work 3 displays the largest negative divergence (-0.470), combining moderate compositional adherence (50th percentile) with high texture divergence (97th percentile).

6 Artists’ evaluation↩︎

We asked artists to evaluate Gemini’s pastiches after their artwork, answering the same three questions about the pastiches produced by ChatGPT in [5], for a direct comparison:

1. To what extent do you recognize your personal artistic language and the coherence of your visual style in this new work? (1 = not at all, 10 = completely)

2. How does this new work inspire you or what thoughts does it provoke? (open answer)

3. To what extent do you consider that the work generated by Gemini has aesthetic or artistic value? (1 = not at all, 10 = very high)

Similarly with ChatGPT’s pastiches, the grades for Gemini’s pastiches varied considerably by artist, as shown in Figure 2. Compared to ChatGPT, Gemini’s pastiches were perceived by the artists as closer to their style, with a mean grade for the first question of 4.99, representing a significant increase of 1.41 points compared to the mean grade obtained by ChatGPT. However, they were perceived as slightly less artistically valuable, with a mean grade for the third question of 4.26, which is 0.67 points lower than the one obtained by ChatGPT’s pastiches. Overall, the artists seem to perceive Gemini’s pastiches as closer to their works, in line with the computational evaluation, which constitutes a weak positive answer to our second RQ.

Answering the open-questions, the artists offered a variety of opinions about the pastiches of their artworks. Most of them regarded the images created by AI more like “artistic comments” than real creations, with little understanding of originality, color mixing techniques, body, and depth. Artistic comments, in this instance, refers to image that functions more as visual interpretation or reactions to the original artworks rather than independent creative works.

We reproduce here some of the answers. Oana Năstăsache says that “some colors look dirty(...) I would never associate that ochre-green combination, because it feels cold and muddy” or that “The surfaces are placed without any sense of composition”. Ion Grigorescu remarks that the new pastiches followed the concept and composition, but failed in spirit, technique, and style: “Young man rising from a chair. Egon Schiele — closer to forgery than to pastiche. A theatricality cheaper than Schiele’s”. Philip Patkowitsch has a different point of view raising the issue of intellectual property: “they actually present my works better, but just because it is using parts of the original image, so I would say copyright infringement would be an issue!”.

So, in the end, did Gemini produce artworks as perceived by the artists themselves? Most of them responded negatively, motivating that AI displayed insufficient command of color theory, which resulted in a poor chromatic atmosphere and the lack of full conceptual comprehension/integration.

7 Interpretation of the results↩︎

The results confirm that artistic style should be evaluated multidimensionally [5]. Both GPT and Gemini successfully match texture statistics (low AdaIN distances) but show substantial compositional divergence (high DINOv2 and VGG19 distances), showing that these dimensions are orthogonal: high texture adherence often pairs with high compositional divergence, and vice versa, as illustrated in Figures 7, 8, and 12 from the Appendix. Thus, this “compositional gap” is an intrinsic challenge for AI pastiche generation, as visually shown in Figure 11 from the Appendix, confirming that visual models consistently extract stylistic features across different generative AI models, positively respondig to our third RQ.

The models also exhibit distinct generation strategies, pointing to a weak positive answer to oure first RQ:

• Gemini: An “interpretive” approach prioritizing semantic alignment (15% closer to CLIP) and higher diversity (+10%).

• GPT: A “faithful reproduction” approach prioritizing texture matching and consistency.

Because of the weak correlation (\(r = 0.293\)) between texture and composition, evaluating AI art requires a multi-metric framework:

• AdaIN-Style: Texture/color fidelity.

• CLIP-ViT-L: Semantic/conceptual alignment.

• DINOv2: Compositional structure.

8 Empirical Remarks↩︎

We observed that Gemini often disregards prompt guidelines, defaulting to the styles of iconic masterpieces (e.g., Egon Schiele or Grant Wood’s American Gothic). Although this exploratory approach allows for originality, the model often overcompensates, producing works that compete with canonical art rather than complementing the original. Both models exhibit perfectionism, frequently completing works that artists intentionally left unfinished, while significantly altering the original image’s foreground.

While Gemini “understands” dimensionality and brushwork, it consistently fails to perceive depth, context, and conceptual irony. A prime example is the pastiche of Ciprian Mureșan’s Pioneers. Mureșan depicted communist youth as rebels and villains to mock 1980s Romanian propaganda. However, both Gemini and ChatGPT missed this irony, independently generating similar images of pioneers as standard patriotic drummers, shown in Figure 9. The striking similarity between the drummers generated by the two models suggests that they were trained on the same visual data.

Finally, Gemini performs better when creating collages and installations compared to paintings. This combinatorial technique aligns with LLM capabilities, though the results often resemble over-accumulation rather than originality.

9 Limitations and Future Work↩︎

Our study is constrained by a small conceptually homogeneous dataset, as most artists belong to the same gallery. Including detailed figurative artwork could provide better insight into LLM pastiche performance. Furthermore, comparing outputs across fundamentally different mediums (e.g., 2D paintings versus 3D conceptual installations) presents structural methodological challenges. Finally, photography was excluded because it functions as a mechanical reproduction of reality rather than an artwork.

We plan to adopt alternative metrics beyond the cosine distance that are better suited for evaluating visual similarity. To mitigate personal artist bias, we aim to incorporate formal evaluations by professional art critics. Lastly, we will apply this methodology to iconic, frequently pastiched artworks, such as Da Vinci’s “Mona Lisa,” Van Gogh’s “Starry Night,” Grant Wood’s “American Gothic,” and Andy Warhol’s “Campbell’s Soup Cans”.

10 Ethical Implications↩︎

This publication raises no ethical concerns, as all software licenses and artist permissions were strictly respected. However, broader generative AI (GenAI) art introduces significant socio-ecological challenges [18]. Societally, GenAI disrupts creative labor markets and raises issues regarding data exploitation, authorship, and intellectual property rights [19], often reinforcing existing inequalities¹. These disruptive impacts on cultural heritage and sustainability require a commitment to responsible inclusive cultural development [20].

Acknowledgments

Research supported by a grant of the Ministry of Research, Innovation and Digitization, CNCS - UEFISCDI, project SIROLA, number PN-IV-P1-PCE-2023-1701, within PNCDI IV, and by UB project.

We are grateful to all the artists who agreed to let us use their works and for their insightful feedback: Adi Matei, Ciprian Mureșan, Ion Grigorescu, Iulia Uță, Karine Fauchard, Lazar Lyutakov, Marius Tănăsescu, Mathias Poeschl, Oana Năstăsache, Philip Patkowitsch, Răzvan Botiș, and Tom Chamberlain.

References↩︎

---

1. https://www.weforum.org/stories/2025/01/the-impact-of-genai-on-the-creative-industries/↩︎