MIN_BILD_AI (2024-2026)
Progressive GAN, 47M params, 7843 images, GTX 1660 Super
Self-trained from scratch on my whole camera roll, photographs of my life rather than curated pictures of work. Two attractors: dark dense forms and bright flat compositions. The failure became a portrait. 858 Python files, 12 repos.
Training progression: first noise to refined
Epoch 1
Pure noise. Nothing learned yet.
Epoch 25
Light and dark begin to separate.
Form
Dense forms emerge, the first attractor.
Progressive
Resolution grows layer by layer.
Detail
Texture and edges resolve.
Full res
The dark dense attractor holds.
Epoch 60
Refining, composition settles.
Epoch 80
The bright flat attractor.
Resolution ladder: 4x4 to 256x256
4x4
The seed grid.
16x16
Blocks of tone.
64 px
Forms cohere.
128 px
Detail arrives.
256 px
Final resolution.
The GAN in motion
Gallery: outputs across the training
