Research Paper :
A Look into Taxidermy as a Synthetic Media
Ideal Animal Image or Trash Data
The practice of capturing animals as visual images, observing them, and recording them reveals the relationship between humans and non-humans. This relationship shows how humans have used or misused the existence and image of animals over time, and what cultural intentions have been inscribed upon them. In particular, capturing the unpredictable movements of wild animals in photographs is challenging. The animal images searched on the internet or seen in typical nature documentary photographs rich with information can, in fact, be said to be images that are deeply ingrained with the visual grammar of how humans view wildlife.
In the case of trail cameras (unmanned cameras) that exclude human vision, the captured images of wild animals reveal a broader variety of aspects, because these cameras do not (or cannot) capture animals in ideal poses like human photographers do. Instead, they randomly capture unexpected angles or partial views of the animals. As a result, many of these photos are considered ‘unusable’ within the framework of human visual standards.
.png)
.png)
Green Korea United’s Trail Camera Archive.
During Unmake Lab’s work process, when experimenting with such photos as a dataset for machine learning, there was a scene where an interesting discovery could be made. When images extracted from trail cameras—including those that were cropped, overly zoomed in, or captured random aspects of wild animals—were used as a dataset and processed through machine learning, the images of animals were predominantly generated as side or frontal views. This is likely because the animal data that existing artificial intelligence has been trained on primarily consists of idealized images within human visual perspectives, such as side or frontal views, and the pattern is learned through transfer learning and generated.
.jpg)
General Biology Research Photos, GAN Machine Learning, Unmake Lab, 2023.
Alternatively, trail camera footage often generates data that could be considered ‘trash data’ due to its random poses and unexpected angles, which is why it may have been excluded during the preprocessing stage. Also, in the case of endangered species, it may be due to the limited amount of available data, caused by their physical rarity, that leads to underfitting in the machine learning process. Just like this, the data bias related to race and gender that exists within human culture can also appear in relation to non-human animals in a different way. This is particularly prominent when applied to real-world environments, where the broader and less predictable context of nature often leads to more mismatches.
This process of work and research raised the question of what an ‘animal portrait’ is, and this text is a record of that research, with an emphasis on ‘taxidermy.’ Of course, this is not only about the existing cultural context but also addresses issues such as classification problems and training datasets together, both of which are being repeated in the field of computer science.
Moon Bear’s Crescent Mark or Botched Taxidermy
Contemporary taxidermy, which represents animal portraits, is rooted in the history of animal preservation. However, as seen in the example above, taxidermy takes on different cultural meanings depending on who wants to preserve the animal and why. The custom of taxidermy has been linked to the aspirations of natural scientific research and colonial practices 1, and today it is associated with issues of wildlife conservation and restoration. Yet, at the same time, it still ties into the culture of trophies. Most taxidermied animals, after death, are represented as though they are alive due to their body and facial expressions, capturing the specific characteristics and expressions of their species in a particular condition. Taxidermy is about preserving the animal’s body, but more accurately, the materiality of taxidermy is preserving the skin.
The skin of taxidermied animals embodies socio-cultural meanings that allow us to contemplate the current relationship between humans and non-humans, while also reflecting humanity’s governance over nature and animals. In other words, ‘taxidermy’ exists as an interface between humans and non-humans. We too collected various taxidermied animals for datasets during our process, and a significant number of distorted specimens were found. These botched taxidermies, occupying space in natural history museums or relegated to their storage, made us wonder about the human gaze and intent embedded within them as well.
.png)
Botched Taxidermy of a Bear, Photographed in 2023. This bear taxidermy is not a moon bear, yet it has white fur on its chest like a moon bear. The white fur appears to have been dyed during the taxidermy process, which may be why this specimen is kept in storage at a museum whose name cannot be disclosed.
Particularly, taxidermy and the culture of trophies in the 19th century are deeply rooted in the practices of imperialist colonization. The colonial relationships between India and Britain, as well as Korea and Japan, are well illustrated through trophy hunting photographs of animals. The grammar of human dominance over nature and animals was tied to the logic of colonial governance and rule. The hunted animal (of that time), often idealized in form, was transformed into an object through taxidermy, becoming a trophy and thus integrated into the order of human and power. This practice was not only preserving the animal’s body but also taxidermying the exploitative history of colonization.
.jpg)
.png)
Tiger Hunting in Colonial India during British Rule. (Photographed at the Royal Ontario Museum, Toronto, Canada)
Tiger Hunting in Korea during Japanese Colonial Rule under the Pretext of Predator Control Projects.
Trophy Hunting or Botched Trophies
The question led to research on how such issues might be revealed through collective records (or datasets) in the realm of computer science. It was also attempting an approach to represent animal portraits and taxidermy through the latent space of artificial intelligence, adopting a method akin to ‘synthetic ethnography.’2 This approach not only dealt with ‘lens-based images’ extracted from reality but also brought ‘computation-based images’ together.
To begin, we experimented with how the culture of ‘trophy hunting’ is represented within generative models, as a way to explore the relationship between humans and (wild) animals. This is also a question of how the visual grammar, including animal portraits and cultural practices and forms represented in trophy hunting, is reflected in generative AI models
.jpg)
An Image that ChatGPT-4o Typically Generates based on the Prompt ‘Create a Trophy Hunting Image.’ (Regarding the trophy hunting images, Google’s Gemini, one of the most advanced AIs, refused to generate images with this keyword. Stable Diffusion was excluded from the investigation due to the possibility of retraining as an open-source model. MidJourney, another prominent image generation AI, was also excluded, as its prompts tend to be filtered and applied in a highly designed and cinematic manner.)
Most of the generated images depict a man dressed in an outfit with safari or camouflage patterns, holding a rifle, standing with a proud expression and stance. The background often features a savannah, bathed in the reddish glow of a setting sun. With this romanticized backdrop, the man stands next to a large animal that is posed to appear submissive to humans, having lost its wildness. This posture, which is manipulated as if it is alive despite being dead, closely resembles real trophy hunting photos that can be found through a Google search. In other words, all the visual grammar is similarly reflected.
What is particularly interesting is that, in some cases, the animals were generated in a way that often amplified their majestic qualities or mythologized them. Another consistent pattern in these images was the prominent presence of a rifle, which was often exaggerated in scale, making it stand out. This was both predictable and unsurprising. However, one intriguing aspect was that even when the prompt was adjusted, this pattern (of the rifle) was not easily removed and left strange, lingering awkward traces behind.
In fact, this weighted correlation might not be limited to the rifle alone. It is not much different from many other strong correlations reflecting culturally fixed associations, such as a bride’s wedding dress in wedding photos, a firefighter with their hose, or a police officer with their baton. Therefore, rather than simply ‘revealing AI’s bias’ or discussing ‘ethical issues,’ this can be seen as an intentional play with these weights, or a satirical ‘trophy collection’ of human culture collectively reflected in AI.
We continuously attempted to remove or replace the rifle by inputting various opposing concepts, attitudes, beings, and objects to counterbalance its weighted significance. For example, we tried numerous prompts such as a baby, planting, praying, or a nature observation telescope—ideas that we believed could offset or substitute the weight. This turned out to be a far more unpredictable play than we anticipated, navigating the fine line between datasets, classification, and the cultural archetypes ingrained in human behavior, inducing chaotic tension within machine learning.
.jpg)
Image from Rifle Removal Research, 2024, Unmake Lab.
Despite various attempts, this deeply ingrained pattern often left behind the ghost of the rifle, and GPT repeatedly asserted that ‘no rifle was generated.’ This phenomenon could not easily be dismissed as a typical hallucination arising from an incomplete learning process. While it might be assumed to be a damaged model or one requiring parameter adjustments, we coined it ‘muscle data,’ implying that the pattern is deeply imprinted within the model.
The machine’s obsessive effort to alter the form and function of the rifle according to the prompt transforms the shape and status of the object—the gun—rendering it powerless in its new context. Yet, this was similar to muscle memory in that the existential reality of the gun, etched in human history, still does not disappear.
We appropriated these glitches, which cling to certain machine models, refusing to vanish like ghosts and endlessly reproducing as objects that stand in for human desire and expression, as ‘botched trophies’ or ‘botched spoils’—much like botched taxidermies. Perhaps this is an ironic humor directed at the technologies that have long served human power and violence, or perhaps it reflects a view of the rifle itself as a broken technology.
The images generated by AI are already almost uninteresting. However, when the algorithm, trapped in a particular loop through the play and product of latent space, reveals its obsessions, we find ourselves intrigued by the human history it exposes.
The research, which became a kind of play on anthropocentrism and its ideas, will explore the yet-undefined concept of synthetic ethnography, followed by research on anthropocentrism with a focus on ‘animal portraits.’
-
1
Joanna Page, Professor of Latin American Studies at the University of Cambridge and Director of CRASSH (Centre for Research in the Arts, Social Sciences, and Humanities), states that taxidermy was central to the imperialist, capitalist, and white cultural preservation efforts of museums. “Their more critical or creative approaches have been described as “botched taxidermy,” “rogue taxidermy” or “speculative taxidermy.””, Joanna Page, Decolonial Ecologies, The Reinvention of Natural History in Latin American Art, Open Book Publishers, May 3, 2023, p. 201
( https://books.openbookpublishers.com/10.11647/obp.0339/ch6.xhtml)
2
Gabriele de Seta, et. al., “Synthetic ethnography: Field devices for the qualitative study of generative models”, SocArXiv, July 15, 2023. (https://osf.io/preprints/socarxiv/zvew4).
This research is based on the residency and research project <When Spiders Spin Dust> curated by Kim Junghyun of Korea and Canada, the UKAI Project, and six participating artists.
https://whenspidersspindusk.com
Curator : Kim Jung-hyun
Exhibition cooperation: Ukai Project
Assistant Curator: Kim Ye-ji
Participating artists: Maurice Jones, Sunjoo Lee, Unmake Lab, Erica Jean Lincoln, Francois Quevilon, Hwang Sunjeong
Sponsored by Arts Council Korea, International Arts Joint Fund, International Arts Network Support