From the initial project proposal, my idea had changed slightly. After experimenting with stable diffusion and doing some research on how to use LoRA’s for stylized images, I found that taking a more autobiographical approach would be more achievable and meaningful. An overview of this project, including my goals and intentions, are as follows:
this project is a semi-autobiographical photo essay about how Chinese adoptees construct memory, recount early childhood, and imagine becoming. I aim to problematize the idea of nostalgia, moving past rather positive feelings of ‘fondness’ and ‘knowing’ of kinship and cultural narratives, instead showcasing the ‘unsettling’ and ‘unknown’. I showcase how fragmented, broken, and lost memories that arise from adoptees’ often turbulent beginnings, as well as the disconnect between their pre and post-adoption lives, powerfully shapes how we think about our pasts. Using AI to generate these images, I also ask inquire how emerging technologies can be used re-imagine digital heritage, and how they are rapidly changing how forthcoming generations are documenting adoption.
From the beginning, I knew that I wanted tow work with stable diffusion to create images, though i was not sure how I would end up presenting them, how many there would be, and whether I would do something else with the photos after they had been generated. However I knew that the style of the images was important. The images from my training data all came from a 35mm minolta camera that my parents took to China in 2003. In our preceding trips to China, such as to adopt my brother, we captured those memories on iPhones and digital cameras. In this way, I argue that technology and ways of documentation are so central to shaping the adoption experience, not only for children, but for their families as well. Hence, retaining the coloring, style, quality, and aesthetic of late 90s/early 2000s film photography was quintessential.
Batch of photos from my personal training dataset
Further, after curating a training set, which had over 2000 images, I began to have a better idea of what types of themes and activities I wanted to portray. Namely, I found several photos of my sister and I and 4 other girls from our adoption group sitting on the infamous red couch at the White Swan Hotel in Guangzhou. This couch is extremely symbolic in the Chinese adoptee community. Held in the hotel in which nearly all US-China adoptive families stay during their trip (the US consulate that handles the paperwork is located in Guangzhou), taking a photo on this couch is a symbolic moment that marks the adoptees farewell to their motherland and the onset of a new life in America. Thus, creating a photo to represent adoptee girls on this couch was intentionally included in the final selection.
For the final presentation, it was really a matter of experiment and seeing what types of photos I could generate, before settling on a specific way to present the photos.
After creating my training data, I began to build the LoRA. This was done with the extensive support by my professor and an online resource for using Python in Google CoLab notebooks to train the LoRA, linked here:
https://colab.research.google.com/drive/1PBHD-HtQjwShzAGsbEGH87w0ixz6CZ7Y?usp=drive_link
Because I have never trained a LoRA before, the biggest challenge was going through the code to try and understand how it was working. I did several trial runs on small batches of images, playing with the parameters, to figuring out the best way to upload the data. I found that uploading my images locally within the program resulted in the least problems and errors, even though I know this isn’t the most practical way.