Drawings I had no memory of making. A machine trained to remember them for me.
This work began with a physical and emotional encounter: cleaning my grandparents' house and discovering my childhood drawings — images I had no conscious memory of making. These drawings, created when I was living inside a war zone, revealed a perceptual world shaped by instability and contradiction. The cheerful colours and naive forms did not match the conditions under which they were made. They were, I came to understand, a child's encoding of survival — a visual grammar formed before language could carry what the body knew.
I created two LoRA models: one trained on my personal collection of photographs across my life, and one trained on the rediscovered childhood drawings. Feeding these models into one another using prompts drawn from the 0wl dataset — grandmother's letters and RNN-generated poetic fragments — I generated new images that merge memory, displacement, and the visual logic of a child who did not yet know she was living through something that would stay with her. The outputs are not illustrations of memory. They are its algorithmic return: the machine reconstructing, from fragments of a life, something that neither the drawings nor the photographs could produce alone.
This work is not treated as a finished artwork but as research in process. The LoRA strength ratios are adjusted continuously — pushing toward the drawings' colour palette or pulling toward photographic realism, finding the uncanny space between them. The algorithm reconfigures impossible childhood scenes, exposing how perception was shaped by contradiction and how memory is not stored but generated.
This work raises the central ethical question of the practice: what does it mean to feed biographical material into a machine, and who is responsible for what the machine produces? The use of childhood imagery — particularly imagery formed during wartime — requires care. I have chosen not to exhibit this work publicly in its current form, recognising that such material can be mobilised to evoke pity or political sentiment without generating genuine insight. The work continues as research: its findings feed back into 0wl and the broader methodology without becoming spectacle.