My friend and I have tried all kinds of hacks to bring personalized images into Oneiras.com — training LoRAs, doing face replacements, you name it. The results were never quite good enough, and the cost + infrastructure overhead made it feel like a distraction.

But last week, Google released the nano-banana model, and it finally does what I’ve wanted for months. Combine that with a three-day weekend, and it’s live: you can now upload multiple images to your profile and generate dream visuals with you in them.

​Here’s an example dream I had.

---

Why multiple images matter

Nano-banana determines your “likeness” based on reference images. If you only give it a single front-facing selfie, it doesn’t have enough context. The result usually looks like someone who could be you… in a parallel universe.

When I gave it:

• Front face
• Both sides of the face
• Back of the head
• Full-body shot (or at least upper body)

…the resemblance was much stronger.

That said, there’s no guarantee. Sometimes one image is enough, and sometimes even five don’t quite land. But on average, more images mean a more “me-like” output.

I’ll let you decide which of these was generated with one image vs. five:

​

---

Uploading vs. sending each time

I also wondered if there’d be a cost difference between:

• Uploading images once with Google’s Files API and reusing them
• Attaching images with every request

Turns out: no difference. Token usage is the same either way. The only benefit of uploading first is avoiding the occasional risk of failed uploads mid-request.

What it costs

Google advertises $0.039 per output image, but that’s only part of the story. Each request also includes input tokens. Here’s the breakdown:

• Inputs: $0.30 per 1M tokens (text + images)
• Outputs: $0.039 per image

​Google’s official pricing docs​

For me, each image came out to about $0.041 total. That’s in line with DALL·E 3 for lower-quality generations, and just one cent more than Google’s Imagen 3 (both of which I already use as fallbacks if nano-banana fails).

---

What’s next?

One thing I haven’t tested yet: using reference images of other people. My dreams often feature friends, coworkers, or even random passersby. If nano-banana can handle multiple people, it could make dream illustrations far richer. I’ll keep you posted as I experiment.

If you’re a dreamer, try Oneiras.com for free and see yourself in a different light. At the very least, now you know how to calculate the real cost of a nano-banana request.

---

P.S. A Debugging Frustration

When I upgraded to the latest version of @google/genai, I kept hitting a cryptic error from Google’s API. The fix? Stop the sandbox connection and reconnect so your Lambda picks up the new dependency versions. Simple — but it still took me a couple of tries to figure out.

Cheers!

Evgeny Urubkov (@codevev)

​