Back to Back with a Copy: A Computational Analysis of AI-Generated Visual Contemporary Art Pastiches

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Image Processing, Data Science & Analytics · Depth: Expert, long

Summary

This study computationally analyzes AI-generated visual contemporary art pastiches, comparing the newer Gemini Pro 3 Nano Banana model against the ChatGPT 5 model. Researchers used five computer vision models (AdaIN-Style, ResNet50-Style, CLIP-ViT-L, DINOv2, VGG19) to measure stylistic similarity across texture, color, semantics, composition, and perceptual features using cosine distance. Findings indicate Gemini produced pastiches with improved semantic alignment (15.2% closer) and greater diversity (10.1% higher average distance between pastiches) than ChatGPT. However, Gemini was slightly less performant on shallow features like color, texture, and perceptual adherence. The study confirms artistic style is multidimensional, with a "compositional gap" where texture fidelity does not predict compositional accuracy. Human evaluations by the original 12 artists perceived Gemini's pastiches as closer to their style (mean grade 4.99 vs. 3.58 for ChatGPT) but slightly less artistically valuable (mean grade 4.26 vs. 4.93).

Key takeaway

For AI scientists and research scientists evaluating generative art models, recognize that a multi-metric framework is crucial for assessing artistic style. Your current evaluation methods might overlook the "compositional gap" where models excel at texture but fail at conceptual depth. Consider integrating diverse computer vision models like CLIP-ViT-L for semantics and DINOv2 for composition, alongside traditional texture metrics, to gain a comprehensive understanding of model performance and guide future development towards more nuanced artistic generation.

Key insights

Artistic style evaluation requires a multidimensional approach, as AI pastiches show a "compositional gap" between texture and semantics.

Principles

• Artistic style is inherently multidimensional.
• Texture fidelity does not predict compositional accuracy.
• Generative models exhibit distinct pastiche strategies.

Method

Stylistic similarity is measured using five computer vision models (AdaIN-Style, ResNet50-Style, CLIP-ViT-L, DINOv2, VGG19) to extract high-dimensional embeddings, then computing pairwise cosine distances between original artworks and AI pastiches.

In practice

• Use AdaIN-Style for texture/color fidelity.
• Employ CLIP-ViT-L for semantic alignment.
• Apply DINOv2 for compositional structure.

Topics

• Generative AI
• Computational Creativity
• Art Pastiches
• Gemini Pro 3 Nano Banana
• Computer Vision Models
• Artistic Style Evaluation
• Multidimensional Embeddings

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• Back to Back with a Copy: A Computational Analysis of AI-Generated Visual Contemporary Art Pastiches — Artistic style evaluation for AI-generated pastiches is multidimensional, with newer models showing semantic improvements but shallow feature challenges.
• Back to Back with a Copy: A Computational Analysis of AI-Generated Visual Contemporary Art Pastiches — Newer generative models improve semantic alignment and diversity in art pastiches but struggle with shallow features like color and texture.
• Evaluating Generative AI: Did Astral Codex Ten Win His Bet on AI Progress? — Human evaluation reveals current image generation models still struggle with complex compositional prompts, despite scaling improvements.
• Visual Retrieval-Augmented Generation for Silhouette-Guided Animal Art — Visual-RAG computationally interprets ambiguous shapes, generating animal art from natural silhouettes using retrieval and diffusion models.
• ArtMine: Discovering and Formalizing Artistic Processes — ArtMine formalizes artistic processes from historical evidence using abductive reasoning and self-reflection, shifting AI from artifact generation to process modeling.
• A Zipfian Analysis of Visual Token Distributions for AI-Generated Images — AI-generated images statistically diverge from natural visuals, especially with complex prompts, revealing model-specific "fingerprints."

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.