The Iconicity of the Generated Image

Nanne van Noord, Noa GarciaarXiv 2025 · DOI · arXiv

Introduction

Generative AI models like Stable Diffusion are trained on massive, web-scraped datasets such as LAION. These datasets inevitably contain millions of copies of "iconic images"—photographs like Dorothea Lange’s Migrant Mother or Steve McCurry’s Afghan Girl that have become part of global collective memory. This research investigates the "iconicity bias": whether this overrepresentation leads models to default to these specific compositions, or if systemic biases and decontextualized training wash out their historical significance.

Methodology

The study analyzed 26 discrete iconic images using Stable Diffusion (SD) versions 1.4 and 3.5. The researchers employed a multi-faceted approach:

  1. Data Attribution: Quantifying the influence of specific training images on generated outputs.
  2. Computational Similarity: Using CLIP embeddings to measure how closely outputs matched the original icons versus how much they varied from each other.
  3. User Study: Participants evaluated generated images for historicity (identifying the event), theme, and recognizability (identifying the specific source photo).

Testing involved three prompt types for each icon: the name, a templated prompt (e.g., "A photograph of [Name]"), and a detailed description of the visual scene.

Comparison of iconic images and generations across different models and prompt types.

Key Findings: Bias vs. Iconicity

A primary discovery was that systemic stereotypes often override the "memory" of iconic images. For instance, despite the Migrant Mother original featuring a white woman, prompts for "migrant" consistently produced images of people of color in SD 1.4 and 3.5. The model's learned social stereotype for the word "migrant" proved stronger than its exposure to the most famous photograph of that subject.

Furthermore, the study identified a "Description Paradox." While some icons like Guerrillero Heroico (Che Guevara) required detailed descriptions to achieve a recognizable likeness, others suffered. For the Afghan Girl, adding descriptive detail often led the model to change the subject's gender to male, as it defaulted to masculine figures when gender wasn't explicitly stated.

Variations of generated images for the "Afghan Girl" icon in SD 1.4, showing the gender shift in the description-based prompt (right).

Results and Decontextualization

The user study revealed a significant gap between AI generation and human recognition:

  • Low Recognizability: 75% of generated images were not recognized by participants as being based on an iconic original.
  • Historicity Gap: Participants correctly identified the historical context in only 55.8% of cases. While simple concepts like "man on the moon" were successful, complex historical moments often failed to translate.

These results suggest that AI models engage in "acontextual" learning. They treat iconic images as pixels devoid of significance, capturing superficial aesthetics while missing the historical "soul" or specific details that make the image iconic to humans.

Conclusion

While iconic images are overrepresented in training data, they do not necessarily dominate AI outputs. Instead, generative models often prioritize broader linguistic stereotypes and generic visual patterns over specific historical anchors. For AI, these cultural touchstones are often just more noise in the data, highlighting a fundamental difference in how machines and humans "remember" our shared visual history.

Summary updated

Abstract

How humans interpret and produce images is influenced by the images we have been exposed to. Similarly, visual generative AI models are exposed to many training images and learn to generate new images based on this. Given the importance of iconic images in human visual communication, as they are widely seen, reproduced, and used as inspiration, we may expect that they may similarly have a proportionally large influence within the generative AI process. In this work we explore this question through a three-part analysis, involving data attribution, semantic similarity analysis, and a user-study. Our findings indicate that iconic images do not have an obvious influence on the generative process, and that for many icons it is challenging to reproduce an image which resembles it closely. This highlights an important difference in how humans and visual generative AI models draw on and learn from prior visual communication.
BibTeX
@misc{noord2025iconicity,
  author = {Nanne van Noord and Noa Garcia},
  title = {{The Iconicity of the Generated Image}},
  year = {2025},
  eprint = {2509.16473},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.16473},
}
Figure