Igor Kostin, View of the Chernobyl Reactor on the afternoon of the 26th April 1986
It is only in recent years that denoising has stepped out from the sidelines of photography, becoming once again as fundamental to the discipline as it was during its early days, though taking on a myriad of new forms. Noise has persisted as a recursive trace bearing the marks of a photograph's creation, a nebulous concept that is as much the mathematically precise gaussian noise we picture when we think of noise, as it is the ‘piss tint’ on an AI studio ghibli image (1). It is a form which becomes signal through its subversion of signal.
On the 26th of April 1986, 14 hours after the explosion at the Chernobyl Nuclear Powerplant, Igor Kostin captured one of the first photographs of the incident (2). The photograph was taken from a helicopter flyover of the plant, a distance from the window, at a time where radiation emissions 200 metres above the plant reached 1500 REMS (3). The resulting image exists in a myriad of different qualities online, all consistently grainy, with large chunks of colourful noise surrounding the hastily captured ruins of the plant. This was the only photograph from the helicopter to survive, the rest covered by an ‘impenetrable black layer’, and even this image was degraded to quite an intense degree. In an interview, Kostin later confirmed he had intentionally increased the contrast of the noise in his images, in an effort to highlight the power of the radiation. The motors of Kostin’s camera seized after 20 photographs, unable to handle the intense radiation. Kostin switched to a spare, and the same happened again, this time after only 5.
Yoshio Matsushige’s photographs of the 1945 bombing of Hiroshima, like Kostin’s, bear the material traces of nuclear radiation. Matsushige was quick to photograph the damage from the atomic bomb, capturing his first photograph between 11 and 11:30am on the morning of August 6th 1945, just 3 hours after the attack. Since the darkrooms in the city had been leveled by the blast, Matsushige developed his film in a stream, filled with radioactive contamination, the traces of which left intense material remnants in his film. Matsushige too, found himself unable to push the shutter, paralysed by the horror that engulfed Hiroshima.
Radioactive radiation offers a specific situation within photography, in which noise and signal become inseparable elements of representation. The exaggeration of noise as a result of this radiation is consistent within both analogue and digital photographs, despite being invisible to the naked eye, and marks the picture plane actively breaking down. Radioactive radiation highlights the non-visual properties of photography as a medium which only exists due to a variety of chemical reactions caused by radiation. Radioactive radiation offers an alternative to typical beyond-human-vision forms of photography, such as infrared or ultraviolet, which both use specific technological changes to move beyond visible wavelengths intentionally (4). The depiction of radioactive radiation is a default within cameras, it is something that seeks photographic depiction as a muddied form, as a subversion of the visible and predicted photograph.
In Kostin’s case, the radiation is said to have seeped through the casing of the camera, rather than the lens. In the face of a force of this power, communication methods begin to fall apart, their material conditions of existence resurfacing like old wounds. In the case of Chernobyl, the question of whether Kostin’s images were even his, and the fact that they were edited to highlight the noise, brings to focus the political nature of noise as something that emerged far before modern denoising technology.
In the recent optimization and tokenisation of images, these wounds were called forth once more. NFTs were one form which heralded this change, a form of image so optimised as a sign that the subject is almost a QR code, a unique and fungible form which, in this case, acted to transfer and hold capital. The relationship between signal and noise can be revisited following the idea of images as ‘Operational’, a term first used in Trevor Paglen’s essay ‘Operational Images’, which appeared on eFlux in 2014. Operational images, Paglen describes, are images made by machines, for machines. They are products of the military-industrial pioneering of vision and image making technologies, which, Paglen writes ‘Instead of simply representing things in the world... “do” things in the world.’ (5)
Operational Images are images which circulate invisibly throughout the world, attempting to perform a kind of task between two machines. One, very simple, example of this could be a doorbell camera which determines at what point to alert the owner of intruders. This system consists of two aspects, a camera which takes these images, and a tool which processes them and identifies a threat. In this system, the camera is optimised to be as efficient for the processor as possible, and the processor is optimised to understand the camera's images as well as possible. Another example, which only functions due to the tricky nature of defining an image, may be a resumé, which is typically made by an Large Language Model (LLM) (6) and read only by other LLMs. These ‘images’ are often optimised to include hidden white text instructing the LLM directly, while performing the joint task for the company of selecting candidates for interview.
Operational images provide a recontextualisation of signal and noise which, like radiation, shifts signal away from human vision. In this context, signal and noise may be understood as both occurring in the same representative space pictorally (7), while having different values within an operation. Signal thus becomes every element of the image which improves efficiency, and noise becomes everything else. Signal and noise become binary classifications, in which information is either strictly necessary, or strictly unnecessary. The allowed amount of, and classification of noise, thus changes greatly depending on the operation required. Images used for AI training datasets, for example, are typically small, low quality images which could easily be considered ‘noisy’ by humans. If used for the training of diffusion models, these images must be strictly empty of JPEG artifacts, otherwise these artefacts may manifest in the images the diffusion model produces, whether or not these JPEG artefacts are noise therefore becomes dependant on the role of the dataset. Noise in object recognition tasks may be something that the recognition model cannot classify, which may be a seemingly essential part of the human’s intended goal for it, that is nevertheless passed over by the machine. As such, operational images take on a new form of noise and signal, in which the noise of an operational image is dependent only on the machine's understanding of it, regardless of the supervisor's desires.
Photographic technologies have greatly progressed since Paglen first coined this term, and denoising has shifted from a practice to a generative method (8). While early denoising allowed for photographs to exist (9), the actual taking of the photograph provided the necessary generation which denoising was implemented to improve. The widespread adoption of Computational Photography (10) in the 2010s emphasised that this improving of images via denoising was quickly becoming a correction, noted by Hito Steyerl in ‘Proxy Politics: Signal and Noise’. Eventually, though, denoising became a generative method in of itself, and is now the most common form of AI image generation. The models that use denoising, commonly referred to as Diffusion Models begin their generation of images with Gaussian Noise, gradually denoising it into a coherent signal based on training. Early forms of these models reinstated small amounts of noise as denoising progressed, attempting to correct mistakes the model may have made early on. As this denoising began to have widespread political implications, it became necessary to question ‘Who actually gets to define what noise is?’
Through an understanding of images as operational forms, one may accept that Paglen’s definition, one that highlights the ‘by machines for machines’ aspect of operational images, may be slightly flawed. While machinic classification may have objectified this operationality, images have always acted. The photograph holds power simply by being referred to as a photograph, a power that exists through the association of concepts like truth and proof that are granted to photography by most viewers. These weighty associations lag behind modern technology, however they do allow for an expanded understanding of operational images.
From a traditional standpoint a laser is a photograph, a photograph which may be a weapon. photographic power may be understood through the blasting of a drone out the sky, a war photograph becomes both the ruined drone, and the fibre optic cable that allows it to fly (11). These forms are no less detached from photography than Italian Brainrot, which Hito Steyerl describes as an Operational Image spreading Fascist and Futurist propaganda (12). The Operational Image thus moves away from the completion of a task, and towards instead a streamlining of photography’s communicative properties. Photographs, quite literally, become the transfer of heat, light, and information, from one place to another, reduced to their basic modes of transmission. The photograph is the circuit board in the camera, it is the laser which fixes an eye, it is the photograph of the fibre optic cables, but also the force that once flowed through them is the shadow of the body on the wall in Hiroshima, it is the bomb itself.
Endnotes:
1: This refers to the brownish tinge visible on AI generated Studio Ghibli images but also other illustrations generated by AI, the term is typically used on social media.
2: The first photograph of Chernobyl was taken by Anatoly Razzkazov less than 14 hours after the explosion. Razzkazov suffered throughout his life from his radiation exposure that day, and his image shows a plume of white smoke emerging from the reactor. In an interview with his widow, Galina Mikhailovna Rasskazova, she claims that Kostin did not visit the plant until the 10th of May, and also took credit for a number of Razzkazov’s photographs, which were also censored by the KGB. This rumour is hard to confirm, and it is plausible to assume the smoke simply looked different from different angles.
3: Approximately 1710 Roentgen.
4: This idea is slightly simplified here, but the creation of and removal of the low pass filter are specific actions, which while necessary to digital photography are not necessary in analogue.
5: Trevor Paglen, Operational Images
6: Such as Chat GPT
7: That is to say, there is no implication that noise refers here to limitations in technology causing a fuzzy, inaccurate, or low quality image, or even that noise is a distinguishable aspect of the image by a human.
8: Denoising AI image generation models, typically referred to as diffusion models, emerged around 2020, they are now the primary type model used for AI image generation.
9: That is to say, without a desire for any signal to be isolated from noise one could argue that we never would have developed the chemical processes which allowed for photography.
10: Commonly understood as the development of digital cameras to include algorithmic denoising of images in order to improve their quality
11: Here, I am referring to examples such as Israel’s ‘Iron Beam’ Laser defense weapon (though many other laser weapons exist), alongside the use of fibre optic cables to avoid drone jammers in Ukraine, leading to a web of fibre optic waste across the front. As noted in Part 1: Transfer, fibre optic cables typically use lasers for communication.
12: Hito Steyerl: ‘Art in the Age of Authoritarian Chatbots’ at the Tehran Summit (36:00).