Finally! Image generation is actually usable in the classroom

Meet "Nano Banana", the new cuckoo model by Google that everybody's crazy about!

Several times a week, a host of clickbait YouTube channels attempt—and apparently succeed—in making their bread with titles like:

“OpenAI Sam Altman SHOCKS Industry”

“Apple’s New AI SHOCKS The Industry”

“NVIDIA's NEW AI 'Text To Video Takes the Industry By STORM!”

“Claude 3.5's New AI Agents Are GAME CHANGING”

etc.

These guys have run out of superlatives early in 2023 and are recycling the same kind of headlines over and over. In their world, there’s a game changer every other day. I avoid this type of AI news and prefer a quality channel like AI Explained, which does an incredible job keeping AI nerds like me on top of what’s going on while keeping our feet firmly on the ground.

That said, something did come up that does qualify as a genuine game changer. A couple of weeks ago, Google released their brand-new image generation model, aptly called Nano Banana. “Nano” because it’s a lightweight model compared to Imagen or ChatGPT Image, and therefore much quicker, and “Banana” because… well, it’s amazing.

Made with Nano Banana

If you’ve dabbled with diffusion models like DALL·E or Stable Diffusion, I bet that—like me—once you were over the novelty of generating, on demand, a picture of a dog and a unicorn playing poker on the moon in the style of Andy Warhol, you quickly realized how difficult it was to get the AI to match a very specific vision. Even worse, any edit you wanted, no matter how small, would often result in an entirely new image. And forget about character consistency.

To be fair, Stable Diffusion, MidJourney, and some other tools have been able to produce character consistency—after all, AI influencers on Instagram have been around for a while now. But it wasn’t that easy to pull off, definitely not something you’d attempt in a classroom, and it definitely wasn’t free.

A big leap came last year when OpenAI updated image generation within ChatGPT. For the first time, we had a free model that was much better at respecting prompts (prompt adherence) and could even include cool ( and intelligible) text inside images. Then Google released Imagen 2, which was arguably an incremental improvement over ChatGPT Image. Still, editing capabilities and character consistency weren’t really there yet, which made the educational applications of these tools rather limited. But now, the game has truly changed.

Nano Banana is available for free within Gemini, and even though some people say the purely text-to-image side isn’t mind-blowing, its editing capabilities and character consistency are remarkable. For the first time, we can use these tools to easily craft narratives, build characters, and explore a whole host of new creative ideas.

Here’s an example to illustrate what I mean: I gave Nano Banana a picture of myself and then requested a bunch of edits—changing the photo angle, my clothes, my mood, my age, even the entire setting. I made a GIF with these results. Once again, only the very first photo with the label is real; that’s all the model had to work with.

input: first photo only and various text prompts.

Isn’t that cool? I also gave it a photo of myself and my son posing in front of a medieval armor display in Italy, then had it generate a scene where we’re fighting a knight in the Colosseum. Nano Banana sticks to your prompts really well and keeps whatever elements you want to preserve. I’ve also seen plenty of examples on YouTube where people give the model several different photos and ask it to extract particular objects and recombine them into a new image.

Even more fun: I created a character and gave it life using Nano Banana. I drew a simple picture, uploaded it, told Gemini his name was Stéphane, and explained that we were about to create a story for him. I worked at it for a little over a half hour—frustrations included, because sometimes the model just wouldn’t comply. The following GIF shows my original drawing, then Gemini’s version of the character, followed by a few scenes. I had to stop there to pick up my kids, but even without text, you can see the beginnings of a story.

First drawing is from me, the rest is by Nano Banana under my textual directions.

How could we use this in a classroom?

Up until very recently, I’ll admit that I struggled to find educational uses for image generation models. The only idea I had was with young children or ELL students to practice descriptive language, but the inconsistencies in accuracy were so glaring that I never pushed teachers to try it.

When ChatGPT Image and Imagen came along, the ability to add text and create diagrams looked promising, but the inconsistency was still too high.

Now, though, this feels different. Nano Banana opens the door to real classroom uses across subjects, especially English. Students who don’t feel confident in their drawing skills (like me) can now produce interesting comics and stories. And for the students with artistic talent—well, I can’t wait to see what they do with it!

EdLaw 2D and student data privacy

Something big also happened that I’ll talk more about soon: Gemini is now available in Google Workspace for Education accounts. I have two teacher Google accounts—one with my agency and one with a district I work in—and while I couldn’t access Gemini under those accounts last year, I now have access with both. If you’re in a Google school, you should give it a go.

I don’t yet know if students have access. But even if they do, that doesn’t mean we can have them use it right away under data privacy law (EdLaw 2D here in New York). As I explained before in detail, in my article My Long-Term Prediction for AI in Education, it’s only a matter of time before EdLaw 2D–compliant, student-facing services start integrating these tools. The fact that Google released Nano Banana in the free tier of Gemini suggests it won’t be long before we see it in SchoolAI or MagicSchool.

So go have fun with Nano Banana, and start thinking about how you’ll integrate it in ways that let your students be creative.