Artist Fine-Tunes SDXL on 60 Childhood Photos to Turn AI Hallucinations into Memory Reconstruction
Reddit user uisato presented an experimental art project: fine-tuning SDXL on 60 childhood family photos, deliberately avoiding faithful reconstruction and letting the model generate unstable variations—spaces, faces and fragments that feel familiar yet never existed. The work treats generative "hallucination" as an analogy for recall, offering a tangible case of AI image models moving toward conceptual art.
Confirmed
- The project fine-tuned SDXL on 60 childhood family photos, with a toolchain combining Kohya fine-tuning, a TouchDesigner audio-reactive geometry system, and a modified WarpFusion, merging several pipelines into a complete presentation.
- The core idea maps model "hallucination" onto memory: recall is not retrieval or replay of stored images but reconstruction from incomplete traces, echoing the academic view that episodic memory is reconstructive rather than reproductive.
- The author stresses this is not faithful restoration of old photos; the outputs feel familiar yet may never have existed, mirroring the distortion and blur of human recollection.
Why it matters
- The project turns a flaw—generative hallucination—into an artistic medium, making the structural similarity between AI hallucination and human memory distortion intuitively observable.
- Using personal childhood photos as an emotional vehicle, it translates a cognitive-science insight into a watchable, empathetic visual experience, showing the potential of AI image models in conceptual art.
2026-09-03 ~ 2026-09-05 · 5 related posts
Primary sources
- Fine-tuning SDXL on childhood photos to simulate how memory reconstructs the past — Chuka444 · 2026-09-03
- [source] Redditor fine-tunes SDXL on 60 childhood photos to simulate memory recall — uisato · 2026-09-04