Fixion2026
Images

1 / 2

2 / 2
Video
※ Quality: Use player settingsExhibition History
Description
<Fixion> observes where a diffusion model takes a face when asked to redraw it 150 times without a prompt. Iterating 126 demographically balanced faces from FairFace unconditionally through Flux.1-dev img2img, every trajectory reached a limit cycle—a state in which the model endlessly produces slightly different versions. No face was ever completed; the 126 did not merge into one, and no convergence bias by race, gender, or age was observed. Whether this "fairness" is the absence of bias or indifference to difference cannot be settled by measurement alone, and the work leaves that question to the viewer.
Caption
Joonhyung Bae, <Fixion>, 2026, Video, 1 min 15 sec
Credits
Artist: Joonhyung Bae
Press
Fixion — CVPR Art Gallery 2026 (Denver)
- CVPR 2026 Art Program to Spotlight Computer Vision's Influence on ArtCVPR 2026 (Computer Vision Foundation) · 2026-06-05
- Fixion — Joonhyung Bae, KAIST (2026)CVPR AI Art Gallery (thecvf-art.com) · 2026-06-05
- Seeing Machines: Luba Elliott on the 2026 CVPR Art GalleryLe Random · 2026-06-01
Keywords
Video
AI
Generative Model
Bias