On bananas and tourists from the matrix

Last weekend, Barış and I went down to Santa Cruz. I am slowly uncovering that Bay Area has an interesting geography, so an hour in the car is often all it takes to go from snow-bright ridgelines to a beach with glitter on the water.

We found a great coffee spot, and let the ocean do that reset thing it does.

Back home that night I scrolled through our shots while my tech-infested timeline kept feeding me demos and papers. That is when the thought landed. We had driven an hour to change the backdrop. With the newest tools popping out I could be even faster.

I had been playing with Nano Banana (a compact image model that shocked the community, later revealed to be Gemini 2.5 Flash Image). Same hair and earrings, same jacket and little quirks, carried from one scene to the next. It felt like moving a paper cutout version of a person across a stack of postcards.

That tugged a thread back to the StyleGAN era. StyleGAN introduced a style-based generator that separates high-level attributes like identity and pose from finer stochastic detail, which is why you can tweak one without wrecking the other. The original paper spells this out and shows how the architecture improves interpolation and disentanglement across scales.

Around the same time, thispersondoesnotexist turned that research into a cultural moment. Tom Roelandts has a clear explainer of how GANs pit a generator against a discriminator and why that feedback loop can produce such convincing faces, while also noting the little quirks you sometimes spot on a hard refresh.

There is a serious footnote too. Research has shown that identity information can leak from training data into synthetic faces, even without malicious intent. The WACV 2021 paper “This Face Does Not Exist … But It Might Be Yours!” evaluated StyleGAN2 generations with several matchers and found evidence of identity leakage. That reminder sits in the back of my head whenever I tinker.

So I gave myself a small quest. By the powers granted to me by the long-weekend, a generous API access from Google, and Github Student Pack’s DigitalOcean credits loaded on to a 5 USD VPS (and Codex CLI loaded on to it), I rolled up my sleeves.

I spun up the tiny DigitalOcean VPS droplet, pointed a simple pipeline at a batch of synthetic faces, curated a list of famous landmarks, and asked Nano Banana to place a consistent character into each scene. Keep it lightweight. Keep it playful. Avoid repeating the same face with the same place so the tour feels fresh.

With a little bit of back-and-forth, thistouristdoesnotexist was live.

Refresh, and a synthetic traveler appears at a landmark. Refresh again, and another character shows up somewhere new. It feels like a scrapbook from a parallel trip never taken.

A few things stood out while I was play-testing:

• Character identity holds steady from scene to scene. Hair, accessories, and proportions carry over.

• Architecture survives with more fidelity than I expected. Domes, arches, filigree, and skyline edges read as the real location rather than a generic backdrop.

• Poses and light usually make sense. The stance looks like a tourist pose, shadows fall in the right direction, and reflections often line up with the environment.

There are wobbles. A stray hand blends into a railing. A face drifts slightly. I smile when that happens, because it reminds me that these tourists are stitched from matrix multiplications on a TPU, not post-cards from a friend.

The Bay Area makes quick physical travel feel normal. Tools like this make quick imaginary travel feel normal too. I am not trying to replace the feeling of wind on the pier in Santa Cruz. I am sketching with photons and letting a virtual traveler hop from the Hagia Sophia to Chichén Itzá in seconds.

Have fun sending these virtual people on a joyride around the world. If you find a favorite, I would love to see it!