Some time ago, I sat down to make my partner a birthday gift — a photo album that told the story of us. As I sorted through years of saved photos, I noticed something quiet: the moments I remembered most clearly weren't the ones we'd photographed.
The rainy afternoon we wandered around on our first date. A quiet sunrise during a road trip when neither of us reached for our phones. The hours of just sitting together in silence. None of it was captured.
So I tried to make those photos with AI.
That's where I learned something most indie makers working with image models probably already know: most generation tools aren't built for the people who'd benefit from them most. I fired up Stable Diffusion, tried Nano Banana, spent days wrestling with lighting descriptors, negative prompts, style modifiers, weight values, and seed numbers. The tools were powerful, but they were built for designers who enjoy the craft of prompting — not for someone trying to recreate the feeling of a Tuesday afternoon in October.
I wanted to think about the memories. I was spending all my energy on prompt engineering.
So I flipped the model. Instead of asking the user to learn a craft, I figured out the craft myself and put it behind a wall. You describe the memory the way you'd describe it to a friend ("we walked along the beach at sunset and she was wearing the blue dress"), upload a couple of photos so the system knows what you both look like, and it does the rest.
The hardest part wasn't the generation itself. It was identity consistency — making sure that across thirty different scenes, both faces still looked recognizably like the two people in the original photos. Generic image tools handle this poorly because they're not specifically tuned for two-person photography. I spent months fine-tuning prompt engines just for couple-specific cases, including softer everyday scenes and wedding scenes across different cultures .
When my partner finally opened the album, she was amazing and loved the gift.
I figured if it could do that for one person, it might be worth turning into a product.
A few things I picked up along the way that might be useful to anyone else thinking about a small AI product:
1. The interface is the product.
The underlying models are mostly commoditized now. What separates good consumer AI tools from bad ones is the work between "user has a goal" and "model generates output." For couple photos, that meant figuring out things users would never know to ask for: maintaining facial features across multiple generations, getting the proportions of two bodies right, ensuring an embracing pose actually looks like an embrace, lighting that doesn't betray its AI origins.
2. Niches that "don't matter" probably matter to someone.
AI couple photos sounds frivolous next to "AI assistants for code" or "AI agents for productivity." But the people who actually use it aren't looking for productivity. They're trying to fill a hole in the visual memory of a relationship — anniversary gifts, wedding vlogs missing the right shots, long-distance couples who haven't been in the same place in months. Building for emotion turned out to be a real space.
It's at couplegoalsai.com if you want to take a look.
