When I obtained my doctoral degree in 2019, I worked as an image-processing operator, primarily in design, handling many image-processing tasks daily. At that time, the industry was constantly buzzing with the remarkable development of the Diffusion Process initiated by Jascha Sohl-Dickstein and others and the rapid democratization of image processing. *1

​

At the end of 2021, triggered by an experience of phantom pain, I began preparing to restore my accumulated record photos and studying for "editing."*2

In this artwork, 'editing' refers to a technical metaphor for Somatic excavation, in the sense of reintegrating fragmented experiences and memories within a local dataset space (i.e., a personal data repository of the data's feature space). In that sense, this work would not have been possible without such technological development and its open-minded democratization.

​

In Denoise Body Experiments, I begin with photography and then utilize latent diffusion models, several image processing filters (.py scripts) for image processing, and Photoshop. The clothing is based on the attire the model typically wears.

In the reverse diffusion process (reverse SDE: Stochastic Differential Equation, dx = [f(x,t) - g²(t)∇x log pt(x)]dt + g(t)dw), one of the theoretical frameworks of diffusion models, paradoxically learns generation methods by gradually destroying data. This process starts from a complete noise state and gradually reconstructs the original signal while extracting the structure. The generated data reflects results sampled based on probability distributions.

The "generation through gradually destruction" mentioned was essential for "editing" the body transformed by injury. This process initially involved a crude therapeutic method of dissection rather than a reconstruction of the post-injury body. The inseparable connection between destruction and creation links the creative approach to the engineering method. Just as the diffusion model operates on noise, extracting patterns perceived as meaningful structures by humans according to probability distributions that capture underlying data characteristics, this work creates a new wholeness from the fragmented body after injury.
Such "editing" was almost impossible in the physical space in which we live.

As a technical metaphor, the layering and fragmentation of images using latent diffusion models and other techniques correspond to the creator's attitude of "casting the body into latent space as a single grain of noise. "This method enabled the integration of multiple temporal states into a single visual plane and allowed them to be "edited" in digital space. In this way, the creator's sensibility and engineering methods work in harmony without conflict.

​

In these works, I constructed a local dataset comprising approximately 1,000 personal record photographs, which served as training data for the latent diffusion model. The model was implemented using a standard PyTorch framework, employing a U-Net architecture-based environment (a neural network designed to estimate either the mean μθ or the noise ϵθ of the diffusion process). For image processing, I created batch processing filters mainly using OpenCV, performing color control, blur, sharpness processing, etc. *3

NumPy was used to optimize computational processing. I consistently used a command-line interface (CLI) at all stages, practicing impromptu creative coding. This is akin to writing poetry, and while it is neither efficient nor is the code refined, there are creative moments similar to the instant of turning over a sheet of paper freshly printed from a woodcut. 

​

After these processes complete the photographic base, it is printed as a giclée print. Finally, it is completed by hand coloring (dye ink) and drawing before framing. The local dataset is discarded after production.

​

References:

*Note1: 
[1] Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N., & Ganguli, S. (2015). Deep unsupervised learning using nonequilibrium thermodynamics. International Conference on Machine Learning.

[2] Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2021). Score-Based Generative Modeling through Stochastic Differential Equations. ICLR.

Both references retrieved from Google Scholar, August 11, 2024.

​

*Note2:
This phantom pain in 2021 was an extremely physical bodily response stemming from experiences of violence (which I had been pretending never happened) during my childhood through adolescence. I was surprised because I couldn't control it with my own will. This experience provided me with a specific creative objective: the aesthetic sublimation of trauma.

 

*Note3: 
However, to visualize progress, the tqdm library was utilized to track the waiting time during filter processing in the .py script.