Banana science! Using ‘Nano Banana’ (or any image gen) to make visuals your aunt can understand

Imagine showing your latest research to your mother or father, and they actually nod with comprehension. Not just a polite nod, a real I-get-it nod, with real appreciation. Wouldn’t that just be warm and fuzzy? With AI image generation getting to the levels of Google Gemini’s recently launched Nano Banana, snack-sized, easily digestible visuals that can go a long way to make your research more visual, palatable and understandable, are now possible. In this post, we’ll share some tips and tricks on how to weaponise the new wave of image generators to create clear, shareable explainers for non-experts. We’ll also look at its limits, because AI still often thinks physics and scientific accuracy are mere suggestions. Nevertheless, image generation is only getting better. The internet loves to use bananas for scale. Well, now you can use ‘nano-banana’ (Google has lots of competitors) for sharing science simply as well.

Banana what now?

Nano Banana is simply Google’s (admittedly catchy) nickname for Gemini 2.5 Flash Image, a fast model focused on photo editing and text-to-image inside Gemini itself. It can effortlessly blend multiple images, keep a created character consistent across edits, and make precise, object-level changes via natural language. Basically, it has proven exceedingly good (and currently better than its competitors like ChatGPT) at understanding the needs of the prompter, and especially at keeping intact the context within an image you don’t want to change. In other words, it usually succeeds in changing only what needs to be changed in an image.

For instance, Medium.com blogger Thomas Smith took a simple photo of his dog Lance and prompted Nano Banana to put Lance “in Paris surrounded by bananas”:

There are hundreds more examples online. You can use it to turn yourself into an action figure, see yourself in a different decade, or add yourself to your favourite TV show, painting or any famous, bucket list tourist spot. You can even use it to make really cool 3D figurines (and put yourself ‘in the field’ in a fun way, even if you weren’t really there).

Nano Banana, for science!

In the simplest terms, Nano Banana is currently seen as the best AI text-to-image generator, especially when it comes to photo editing. So, what does that mean for your research visuals?

Here’s a useful summary of what Nano Banana suddenly got good at (with other image generators surely quick to follow), and how it might be used by scientists:

1. Surgical photo edits that keep context intact: Swap backgrounds, recolour lab gear, remove distractions, or add a pointer arrow without melting the rest of the image. Reviewers note Nano Banana’s strength is editing existing photos while preserving the scene, not just hallucinating new ones.
   

   

   “Create a realistic photograph of a bearded researcher, wearing sunglasses and protective clothing, taking a selfie 'in the field' at the top of a volcano, with the lava in the mouth of the volcano in the background.”

   

   “Excellent, now add a banana for scale.”

   
   

2. Character and pose consistency for storytelling: Need the same “mascot scientist” in three panels? This used to break easily. Hands-on tests and press coverage point to stronger identity consistency, which helps when turning a study into a mini-comic.
   

   

   “Put the same researcher in a selfie in the Maldives, dressed for a scientific conference.”

   
   

3. Style transfers and reference matching: Apply the colour palette or texture of Image A to Item B. This is perfect for keeping your figures on-brand with your lab or university style guide. Or to pitch the schematic of your anti-gravity propulsion engine in the style of Blade Runner 2049 screenshots.
   

   

   “Recreate the first image in this thread in the style of Blade Runner 2049.”

4. Multi-image compositing: Blend your microscope photo with a clean backdrop and a caption card, or merge a field shot with an annotated layer, all inside one prompt chain.
   

   

   “Now add my uploaded field notes as a sticky note in the corner of the photo.”

   
   

5. Speed and accessibility: Press tests and user reviews emphasise quick generations (an image generation takes ~15 seconds instead of ~60 seconds with ChatGPT) and an approachable interface. Also, it is FREE (it works even on the Gemini base model). That makes it especially viable for researchers who do not want (yet more) software subscription costs in their lives.
   

There are a couple of things to remember: Firstly, although Nano Banana is trending, it won’t be the best forever. Next week, some other AI company (OpenAI, wink wink) will likely come up with something even better. The point is not to get overhyped on any specific tool, but to learn how you can use that tool. Secondly, it is not without its faults, so never trust AI completely, especially when it comes to research and scientific accuracy (more on this below). And lastly, all Gemini images include a visible watermark and an invisible SynthID tag, a technology from Google DeepMind that identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text or video (just something to be aware of, since you will need to credit AI visuals used in research).

Wait…that’s not right

So, where does AI image generation (including Nano Banana) still drop the Newtonian apple when it comes to scientific visuals? Let’s see:

1. Environmental and taxonomy mix-ups: Our head-honcho and marine biologist Tullio Rossi came up with a nifty “oyster reef test” for AI image generators. Until recently, if you asked an AI to create an image of an “oyster reef”, it would serve you up a very scientifically inaccurate seafood platter, a la this one from ChatGPT with open oysters as if they were served at a restaurant:
   

   

   
   

Now, Nano Banana has got it (mostly) right, minus some tropical fish species that clearly don't belong:

The point is, don’t trust AI to know the environment, taxonomy or context of your field of study as well as you do. Scrutinise every detail, and ask for it to correct mistakes, so you don’t end up embarrassed.

2. Physics that does not add up: Watch out for impossible shadows, weird reflections, and improbable object interactions. Industry guides still call out basic physics errors as reliable telltales of poorly generated AI images. For instance, look at the funky physics in this dump-truck example:

A truck dumping a pile of research papers according to Nano Banana

3. Beware anatomy: Anatomical nonsense can easily sneak into slick-looking diagrams. A 2024 Frontiers article was retracted after the now-famous “endowed rat” AI-generated figure with garbled labels and laughably incorrect anatomy slipped through. Ethicists warn that absent strong guidelines, erroneous AI figures will keep popping up. It’s best to build your figures from vetted examples or templates, not from scratch. Otherwise, you generate at your own peril. For example, even with Nano Banana, AI cell biology knowledge is still a disaster (biologists, try not to throw up):
   

   

   A prokaryotic cell according to Nano Banana

   
   

4. Text and labels: AI still struggles with crisp, correct typography in-image (again, see the cell abomination above). To be on the safe side, always ask AI not to include labels or text in scientific images so you can add them yourself after the fact, to ensure accuracy.
   

5. Chemistry: AI does not understand chemical structures! A researcher who understands chemistry better than I do recently flagged this figure that was published in MDPI as Figure 2 of a paper that, as of today, is still not retracted. The image is complete rubbish full of typos and chemical structures that make no sense. It was clearly AI-generated and not caught during peer review. Shame!
   

   

   Chemistry AI gibberish recently published in MDPI (https://doi.org/10.3390/chemosensors13090337)

6. Safety, disclosure, and watermarks: As briefly mentioned before, if you share AI visuals publicly, for a paper, a popular article or social media, keep the visible watermark and declare AI assistance (and the model source). This way, it won’t bite you in the ass later, in terms of copyright or (heaven forbid) when you need to defend some inexplicable inaccuracies.
   

Mini workflow: from idea to post in 12 minutes

To make things easier for you, here’s a nifty little mini workflow to help you brainstorm how to use image generation to give your study some visual pop:

1. Pick one digestible idea from your research (what changes, what causes it, why it matters).

2. Gather 1-2 relevant, useful reference photos or figures for the AI to study.

3. Generate or edit with Nano Banana (or whatever you prefer) until you like what you see, but keep text out of the image. Don’t be scared to iterate and play around!

4. Add arrows, labels, and a short caption manually after the fact.

5. Fact-check against a source and add a citation of evidence (if possible), and attribution as AI-generated (noting the model used).

6. Export, include alt text and relevant missing titels and labels, and you should be ready to post!
   

To summarise, here are some guardrails to remember so your visuals do not backfire:

• Always cite a source for the science in the caption (did you get the data from your own research, or from somewhere else?), even for a schematic.

• Keep watermarks and say “image created or edited with AI” in alt text. Gemini adds a visible and invisible watermark by default.

• Check physics for logic (doh!).

• Never trust AI labels! Add text in a design tool and if possible, peer-review it.

• For biology and medicine, use AI as a sketch generator, then redraw accurately or commission a scientific illustrator.

• Have a domain expert sanity-check the image if it touches on issues related to safety, health, ethics, or ecology.

Go bananas!

With all of that being said, remember that this is a really fun tool and you shouldn’t be afraid to play around with it (just make sure the science is sound in the end).

AI is finally useful for outreach-ready visuals, as long as you keep one foot on the brake. Nano Banana makes edits fast and keeps your character on model, but it still makes silly guesses. Treat it like a bright intern: great at drafts, never your final authority.

Want to go deeper with AI in your research? We have another blog post on how to create eye-popping scientific posters using AI, and our course “AI for Researchers” covers safe, effective prompting and visual workflows.

See you in the lab, or in the field, and hopefully, on your aunt’s Facebook feed!

If you found this blog useful, please consider subscribing to our newsletter!