PERSPECTIVE
Unexpected beauty in errors: AI’s intriguing flaws
10 December 2024 – Vol 2, Issue 4.
When we think of errors, we typically associate them with failure, regarding them negatively as something that signifies a mistake or imperfection. AI-generated images sometimes present erroneous results, with often unexpected combinations of elements, and distorted forms. These errors may not align with the conventional idea of failure or imperfection. Rather than being seen as flaws, they can enhance the work by introducing unexpected outcomes.
I will share these unexpected outcomes, as well as some prompts from my experiments with creating AI images.
My curiosity about AI imagery was piqued when I read an article about the world’s first AI-prompted photo prize. The prize was awarded in October 2023 to Annika Nordenskiöld for her image Twin Sisters in Love[i], which features two women cuddling an octopus. “I understand the fear of AI and find it somewhat healthy, but I see it more like a colleague I am working with” Nordenskiöld said. Her monochrome image features two similar-looking women, both seem to be smiling and are cuddling an octopus and each other. However, it looks as if they may be conjoined, in what may seem a typical AI ‘error’.
The photographer Charlie Engman creates surreal AI work with distorted human forms. The image Pony Couch (Kiss) presents two boys sitting on a sofa with a pony. The boys are slightly merged together and into the pony, who then is merged into the sofa. It has a recognisable AI photo-realistic aesthetic, almost illustrative and full of AI flaws.
AI-generated image. ‘Pony Couch (Kiss)’ by Charlie Engman. 2022. Permission to publish the image was obtained from Charlie Engman.
These works, together with my lifelong experience as a professional photographer, sparked an idea in my mind – to create AI images in the style of documentary photography that will present unlikely encounters between human and animal characters. I also wanted to explore any errors that AI would generate.
The AI images that I created were then printed using high-quality digital archival techniques, used for the printing of photographs at exhibition quality. The works were printed at almost A2 in size (61 x 40 cm) and exhibited. At that size, the AI errors are not obvious at first glance. As the images allude to a certain ‘truth,’ I was intrigued by how people would react to what they see. How would their perceptions shift when they realised the works were created with AI, and how would they respond when they began to notice the subtle errors?
AI-generated image. Park Life series. ‘Maximillian and Shane’, by Dave Gibbons. 2024.
Much of my visual art practice over the years has been about highlighting imperfection. One project, Flat Pop, is a series of drink cans that were crushed by cars. The cans were photographed under controlled lighting conditions. Through using the transformative qualities of light, I hoped to reveal imperfections, such as scuffs and scratches. I think it’s less about recognising errors, and more about flaunting the imperfection which exists when using found materials.
Photograph. Flat Pop series. ‘Reduced 03 recto’, by Dave Gibbons. 2023.
Starting to work with Artificial Intelligence tools (AIT), I initially made many clumsy experiments. Soon, I was drawn to Midjourney, which seems to be regularly producing authentically photographic results. Further research unearthed a community of artists and researchers who are exploring AIT and are open to sharing their prompting process and workflow.
I was introduced to another AIT, Magnific AI which I use to upscale the images. The maximum resolution currently available in Midjourney is 2912 x 1632 pixels, which limits the scale and quality of printing. With Magnific AI I can increase the pixel dimensions and produce professional-quality prints.
Correcting errors
Within the ‘traditional’ photographic practice, an imperfection (error) is often something which is to be corrected. Filters and tools to adjust and ‘improve’ images are included within digital editing workflow. Adjustments to exposure, white balance and colour correction are commonly used and are often saved as presets which can be accessed with a single click.
Corrections are also used in traditional darkroom processes where a print is made by exposing light through a negative film treated in a chemical bath. With this technique, selective exposures can transform elements of the image, for example, to darken the sky (‘burning’) or to lighten an area of the image (‘dodging’). If a speck of dust or a scratch on the negative becomes visible in the resulting print, this can be retouched using pigments or by reduction using the chemical potassium ferricyanide. Retouching is a standard technique to retouch skin blemishes, wrinkles, and spots.
AI errors, on the other hand, offer a very different experience. There is a process of discovery since the errors are not all obvious; many are very nuanced. For example, on close inspection of the Park Life AI images, we find subtle errors. The AI work Gencio and Chichi portrays a man and a llama:
AI-generated image. Park Life series. ‘Gencio and Chichi’, by Dave Gibbons. 2024.
When zoomed in we notice that what is supposed to be grass in a park is in fact a view over a forest:
AI-generated image. Park Life series. A detail close up ‘Gencio and Chichi’ showing a view over a forest, by Dave Gibbons. 2024.
Another example is the crease between the thigh and abdomen of the llama. When inspected closely it looks like a river valley.
AI-generated image. Park Life series. A detail close up ‘Gencio and Chichi’, by Dave Gibbons. 2024.
To be able to notice these nuanced errors we require the skill of reading images. Yet, there is a general tendency (and perhaps laziness) to take images for granted; to simply consume images.
The back-lit wispy hair on a human character, in Evangelina and Arnold, when magnified, looks very similar to the ostrich pairing.
AI-generated image. Park Life series. ‘Evangelina and Arnold’, by Dave Gibbons. 2024.
AI-generated image. Park Life series. A detail close up ‘Evangelina and Arnold’, by Dave Gibbons. 2024.
Text-to-image process
AIT uses text prompts to generate images, while traditional photography involves physically finding and capturing scenes. Yet, the two processes do share some similarities. In both cases, the artist needs a clear vision of what they are looking for and want to create. With AI, the artist describes a scene they imagine, through text prompts, while in photography, the artist seeks out a scene (or stages a scene) that they wish to capture with their camera. In either case, the intention behind the creative process is the same: I should know where I am going and what I am searching for.
However, the AI process of achieving this vision is different, as text prompts can be refined endlessly. Part of this prompting/re-prompting inevitably leads to human error, such as typos or awkward phrasing. These mistakes, or even small changes in the prompt’s structure or wording, can lead to unexpected and varied outcomes. For example, ‘The brown fox jumped over the lazy dog’ would offer a different result from ‘The fox jumped over the dog’, or ‘The brown ox jumped over the lazy dog’. This is an interesting experiment in creativity.
In my project, a base prompt was used and then adapted for each image, and further adjusted as required. A prompt example for the series is as follows:
“a natural looking medium format photograph of a [human character]. [He] is wearing a [clothing]. The [human character] looks confused and has a solemn gaze. The [human character] is sat on a simple park bench with his chin cupped with his left hand and his right elbow is resting on his knee. A [animal character] is sat in front of the bench and slightly to the right. The [animal character] is looking at the [human character] and seems confused. Photographed in the style of [photographic style] on [film stock], [lighting conditions].” Parameters: Style – style raw Aspect Ratio –ar 16:9
Midjourney did not quite offer the results I hoped for, probably because I did not know what I wanted… Of course, I have an idea which is open. The process is very iterative, and there’s a lot of work involved in the prompting, re-prompting, fine-tuning and curating the responses. The term co-creation is often used when AIT’s are utilised within the creative process. In that respect, the AI process is quite similar to how I approach film and digital photography: the process of selecting, refining, curating.
The following prompt resulted in the image Boris and Titian.
“A wide street style photograph of a bald white man wearing an a pink ballerina’s leotard and tutu with ballet slippers, he is sitting on a park bench. A large male swan is standing next to the bench and slightly to the right and is attempting to comfort the man. The swan looks slightly confused, warm light is dappled through the trees, Photographed on Kodak colour film”
AI-generated image. Park Life series. ‘Boris and Titian’, by Dave Gibbons. 2024.
The AI result of the right sleeve of the leotard seems as if it is embroidered. On closer inspection, it looks like a stencil cut into the fabric, resembling a sleeve tattoo which may or may not exist beneath.
AI-generated image. Park Life series. A detail close up ‘Boris and Titian’, by Dave Gibbons. 2024.
I am not sure I should describe these as errors or faults… perhaps there is some kind of algorithmic creativity, which I do not quite understand, but I am certain that I do not wish to correct it. I want it to play out.
At a Glance:
Artwork = Prompt + Flaws (Data limitations, Interpretation errors, Human errors) → Unexpected Beauty.
AI-generated artwork is the result of human input through the prompt, AI’s flaws stemming from data and interpretation errors, and human mistakes. These together can lead to unexpected beauty.
© Journal of Creativity and Inspiration.
Images © the artists, as specified.
Dave Gibbons is a visual artist specialising in photography and collage. His work investigates the emotional connections people have with found objects and ephemera. His photographic projects, often spanning decades, focus on documenting objects in situ or the studio, exploring memory, nostalgia, and mortality. His collage practice incorporates both digital photomontages and papier collé, drawing inspiration from instructional diagrams and everyday materials.
Footnote:
[i] See article at: https://www.forbes.com/sites/lesliekatz/2023/10/08/see-the-surreal-image-that-just-won-an-ai-art-contest/
