.webp)
Creating an AI photobooth with soul
Our journey to build a photobooth with experimental AI technologies that delivers human connections
Adventures In Pharma was the top-rated AI conference this year*, and one reason we created it was to show off all the technologies and ideas we'd been quietly iterating on for client events for years. We’ve been learning what works, what fails spectacularly, and what makes people laugh or say wow.
(*one attendee described it as "the most engaging life sciences conference I have been to in a very longtime – if ever.")
Phase one: the astronaut experiment
In early 2024, when AI image generators were still consistently making people with seven fingers, we ran a space-themed internal training event for a client, and decided to transform the whole team into astronauts using a photobooth. The results were hit-and-miss, but when it worked, people loved it, and you can still see many of these images as avatars on Teams and LinkedIn. The setup was simple. We used an iPad for capture, sent the photos to whichever AI image service was least broken that week, and printed them out on a consumer photo printer that jammed every twentieth print. The code I'd written to manage this workflow was fragile, but it proved the concept.
Phase two: adding movement
In early 2025, the same client wanted us to level up their internal event. So we used a better AI image model, and added video generation to the mix. The aim was subtle animations: a wave, a thumbs up, answering a space telephone, like the moving photos from Harry Potter, but completely on brand.
With this, the technical challenges multiplied. Now we've got image generation, video processing, and trying to display everything on a gallery screen while three printers worked overtime. We overhauled the code to manage this messy workflow: queuing, processing, catching errors, and somehow keeping it all running when the photobooth got very busy at the coffee breaks.
Phase three: Adventures In Pharma
One of my favourite things about Adventures In Pharma (I have many) were the "Adventurers Team" – six distinct characters which formed the basis of the branding. These were created and illustrated by the fantastically talented friend of Camino, Salvador Lavado, who normally applies his talents for the likes of Taylor Swift and Kaiser Chiefs.
Our own design team of Dan and Anna also gave each character a unique environment that was used to signpost the conference sections: jungle, underwater, etc.
With such a strong and distinct visual identity, we had to do the photobooth, but the challenge was making attendees look like they belonged in this universe, rather than generic "astronauts".
So there was a lot of trial and error. No text prompt could capture the nuance of the team, so I supplied the illustrations as examples. Counterintuitively, giving the GPT-4 image model all six characters was too much info, resulting in confused images that barely matched the conference style.
What did work was an iterative approach to build the photo up in layers. We gave AI the attendee's photo and just one randomly selected character illustration. Then we used AI to fix the"Adventures In Pharma" chest logo separately. Lastly, we composed the character onto the correct background (without AI!), before passing to the AI video tool.
Getting there meant testing everything. OpenAI's GPT-4o understood the illustrative style best. Google's"Gemini 2.0 Flash Experimental" made beautiful photorealistic faces but failed at basic instructions. One in three times it just concatenated images side by side like a bad PowerPoint. Runway fell somewhere in between. For video, we started with Runway v3 but landed on Kling v2.1, which had a tendency to add random explosions we didn't ask for… sometimes bugs become features!
And this is where we found the 'soul' for the photobooth. Rather than using AI to create generic images, we started with genuine human artistry and used AI to amplify it, to allow every attendee become part of that world.
The human layer
Technology is of course, only a small part of this story. We have someone on the booth at all times. Not really for technical support (the code is now pretty bulletproof), but for the human moments. Like when AI kept 'correcting' someone's distinctive hairstyle in every generated image, despite our best efforts. We ended up manually editing their images overnight to restore their actual look. These edge cases matter.
We've now got three printers, because each one takes a minute or two, and the photobooth was most popular during the coffee breaks. So we need to churn them out quickly. We added a gallery too, so that attendees can see everyone's transformations accumulating throughout the day. It created a shared experience, not just individual moments.
How we use AI
This photobooth represents our entire approach to AI at Camino. We start with human creativity. Those hand-drawn characters gave the project its soul. Without the human artistry, we'd have generic, forgettable outputs.
Iteration beats perfection. Our first version barely worked. Our latest version still occasionally hiccups. But each iteration teaches us something essential.
We layer intelligently, not trying to make one AI model do everything. Each tool handles what it's best at. This is how we approach content creation too: AI for first drafts and ideation, humans for strategy and nuance.
We maintain oversight at every step. Every single image could be deleted if inappropriate. Every process had a human checkpoint. In pharma communications, this isn't optional, it's essential.
We build for flexibility because models improve monthly. Our workflow can swap in new models without rebuilding everything. Content strategies need the same adaptability.
And we measure what matters. We track engagement, not just technical metrics. A 95% accurate AI application that no one uses is worthless, a 70% accurate application that delights users and drives engagement is worth the trade-off.
What's next
We'll keep iterating on the photobooth. This week we have an internal event for Camino where we're testing out detecting how many people are in the original photo and using pictures and words to choose which medieval character you end up being…
Of course, next week, OpenAI will probably release something that makes this entire technical stack obsolete, and Google will announce a model with an even more confusing name than "Gemini 2.5 Pro Ultra Image Mode 3."
But for all our AI projects, we'll adapt to the changes, and ensure we're still using AI to create delightful and meaningful engagement with human control. Understand your audience. Leverage the right tools for each task. Maintain human oversight. Never stop iterating.
And keep the soul in everything we make.
Want to explore how this approach could transform your next conference or campaign? Let's chat about bringing some AI magic (with human soul) to your next project.
