Tel Aviv

Buildings a machine dreamt of the city, trained on its architecture, sharpened through every new image model and framed as Polaroids with addresses that never existed.

Project
Self-initiated
Fragments of a Future Past
Craft
StyleGAN2 · Runway · Colab
Stable Diffusion · Interface
01

Fragments of a future past

What does a city remember about itself, especially the parts it never built?

When AI was still in its early, unpredictable phase, I set out to explore Tel Aviv through its architectural ghostprints: buildings that never were, dreamt by a model trained on the city.

Each film shows the whole development in one go: the same building as the StyleGAN2 dream, its Stable Diffusion XL reading and, last, a pass through Krea's enhancer, the furthest stage the project reached.

I stopped there on purpose. At some point the buildings were realistic enough, and running them through every newer model would have made them sharper but more generic, losing the character that made them feel remembered.

The project wasn't about realism. It was about resonance.

One dreamt building in three versions, side by side
Fragments 01one building, three models
One dreamt building in three versions, side by side
Fragments 02one building, three models
One dreamt building in three versions, side by side
Fragments 03one building, three models
One dreamt building in three versions, side by side
Fragments 04one building, three models
02

Work in progress

I started in Runway, training a custom StyleGAN2 model on photographs of Tel Aviv's architecture. After 24 hours the results were uncanny: odd, unfamiliar structures that still carried the character of the city.

I moved the model into a Colab notebook by @dvsmethid and kept going: another 94 hours of machine dreaming. The AI learned not just to mimic the architecture, but to take buildings apart and build new ones. They lived somewhere between data and daydream. But they felt remembered.

With time, new tools emerged. I fed these ghost buildings into each new version of Stable Diffusion. Every model sharpened the image, and laid its own strange bias over it. I wasn't chasing accuracy. I was collecting echoes.

It was never Tel Aviv. And yet, somehow, it always was.

An interpolation loop through the first Runway model
A building from the first Runway model
01 Runwaya custom StyleGAN2, 24 hours in
A latent walk through the StyleGAN2 model
A StyleGAN2 frame
A StyleGAN2 frame
02 StyleGAN2 in Colab94 more hours, a latent walk
The ghost buildings through Stable Diffusion XL, slowed down
A ghost building through Stable Diffusion XL
A ghost building through Stable Diffusion XL
A ghost building through Stable Diffusion XL
03 Stable Diffusion XLthe ghost buildings, sharpened
03

The archive

Finally, I built a simple interface that frames each image like a Polaroid: a moment snatched from the margins of imagination. Every picture gets a fake address; every frame holds a piece of architecture that never existed, but somehow still feels like home.

What emerged is a machine-made archive of a parallel Tel Aviv: a sunlit memory of a city that exists only in fragments of code and flashes of longing.

Back to Top

Interest Areas

  • Generative AI
  • StyleGAN2
  • Stable Diffusion
  • Architecture
  • Interface design
The generator's control panel: image picker, growth rule, quadtree, playback and export
The generatorruns in the browser
A Polaroid of a dreamt building, revealed cell by cell
Loop 01four buildings, exported from the generator
A Polaroid of a dreamt building, revealed cell by cell
Loop 02four buildings, exported from the generator