A new CREATe working paper, published in the Journal of Visual Art Practice applies a novel methodology to assess the process of making images using generative AI. The authors, Dr Jim Brogden (FRSA, University of Leeds) and Prof Kristofer Erickson (CREATe) attempted to ‘re-photograph’ four iconic historical photographs, revealing challenges facing creators seeking to use generative AI to produce visual art.
Rephotography is an established photographic practice, involving repeat capture of a scene from the same viewpoint as an earlier image. Rephotography has been used as a measurement tool in the study of environmental or social change, to capture longitudinal differences in a single place. The authors adopted this method to draw attention to the differences in creative practice that emerge by attempting to recreate the same scene using a new technology.
The authors’ attempt to recreate Dorothea Lange’s Migrant Mother (1936) with progressive attempts shown in panels A-D.
The example above demonstrates some of the challenges of using Midjourney. As we see in panels A and B, the system prefers glamorous subjects, out of place in the context of dustbowl America in the 1930s. The authors attempted to get closer to the specific clothes and setting of the original photograph by adding specific prompt instructions in panel C:
black and white photograph of a pensive rural mother evicted from farm, looking to the left with hand on cheek, two young children hiding their faces away from camera behind her back, dirty faces, gingham shirt, tattered clothing, lines in forehead, worried expression, grainy, –ar 4:5.
Since Dorothea Lange was a highly celebrated and important photographer, the authors wondered if Midjourney would be able to identify her style and apply it to the output. When asked it to produce the image ‘in the style of Dorothea Lange’ the results are shown in panel D. There are unmistakable similarities to the original image in the pose, clothing and expression of the main subject.
By invoking the command ‘in the style of … ’ the authors found a shortcut to achieving greater verisimilitude to well-known images that are likely to be reproduced in the training data. This type of command works with other notable photographers’ styles, such as Annie Leibovitz and Stephen Shore. At the time of the research, it did not work with the images by Vivian Maier and Gregory Crewdson (two other iconic photographers featured in the study), suggesting that their work had avoided inclusion in sufficient quantity to influence the AI system.
The full paper is available in the CREATe Working Paper series.
Click on image to download.
On the process of image creation with artificial intelligence: Rephotography using Midjourney AI
Jim Brogden and Kris Erickson
CREATe Working Paper 2025/7
Abstract
When we make images using text-to-image AI systems like Midjourney, what precisely are we making? In the context of significant cultural, political and legal uncertainty about the status of AI-generated imagery, this paper reports on a practical experiment in rephotography. The authors attempted to recreate four notable American photographs in Midjourney: Vivian Maier’s Untitled self-portrait (circa. 1954), Gregory Crewdson’s Untitled (Twilight), (1998-2002), Dorothea Lange’s Migrant Mother, Nipomo, California (1936), and Joe Rosenthal’s Flag Raising on Iwo Jima (1945). The practice of rephotography, which seeks on one hand to faithfully reproduce and on the other hand to reveal difference, enables us to compare human and AI-generated photography as practice. We find that the AI system exhibits bias in its default perspective, likely an artefact of training data scraped from the open Internet. Specifically, we find: (1) AI images give prominence to main subjects in a way that appears to reflect a late capitalist, consumer gaze; (2) AI images ‘hallucinate’ details that reduce human intentionality and control; (3) AI images are detached from the social and historical context that gives human photographs meaning. Despite these striking differences, we consider emancipatory and critical potential for AI imagery as an artform. Creative possibilities present themselves in rendering visible the probabilistic links between datasets, user prompts and outputs.
Full paper can be downloaded here.