I've been having toe cramps lately. So the timing worked out great when Google released the new Gemini image model, codenamed nano-banana. So put on some Hollaback Girl, grab a pillow, and let's make some pics and giggles!
Prerequisites
- A Gemini API key (from Google AI Studio)
- We will use the python Google Gen AI SDK. And again, using
uvfor python project. - In any folder, initialize the project and add dependencies:
> uv init
> uv add google-genai pillowThen we can follow the simple instruction to try out nano-banana's capabilities.
Image Editing
I wonder how many fictional bananas was generated by Gemini over the last few days following the developer guide. Well, we are contributing. Starting from an input image, it's extremely easy to edit the scene.
Image Editing with nano-banana. As demonstrated, nano-banana can generate both a small faceless banana, as well as a much larger banana of unrealistic beauty expectation. The dog's excitement appears correlated with the size of the banana.
Merch Generation
Who doesn't love stickers? With Jack the corgi again serving as the starting template, we asked Gemini to make a sticker, a plushie, and a 3D figurine.
Sticker, plushie, and a figurine of a corgi generated by nano-banana.
Iterative Editing
A super cool functionality of nano-banana is that it can be used as part of a chat (chat mode), making iteratively refining the image very easy. So, starting from the corgi figurine, we asked nano-banana to rotate it left and right, and chained the results into a gif.
Monster Bake-Off
It's really quite wild how fast AI is moving in the last few years. So I looked up what I did for "genuary" back in January 2024. At that time, I was using SDXL from Hugging Face. And the theme for January 24, 2024 was "Impossible objects (undecided geometry)." So I made a bunch of impossible monsters with Stable Diffusion, running on GPU on Google Colab. To test how well the image model could handle "impossible" and "unrealistic" instructions, the prompts were intentionally difficult to satisfy (I thought). Now, reusing the same prompts, let's see how nano banana compares.
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "ghibli style, a fantasy creature with 7 eyes, 3 legs, 2 wings, and 6 antennas, muted colors, detailed." Three legs is too impossible even for image models.
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "pixar style, a monster creature with 3 eyes, 2 legs, hairy, friendly, playing a yo-yo, vibrant colors, detailed." Overall nano-banana's prompt adherence is a bit better (e.g., yo-yo). But again, 3 eyes is where image models draw the line? 😂
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "a cute pokemon with 3 big eyes, no leg, round shaped, flying, in attack pose, soothing colors"
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "an ancient monster, no eyes, fish shaped, big mouth, two arms, sketched, leonardo da vinci drawing style, pencil on parchment paper"
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "digital art, midjourney style, colorful, fantasy creature, giant, four long arms, two legs, horn on the head, detailed, 8k." So far, nano-banana is a lot more proactive in fully rendering the backgrounds.
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "an impossible object with undecided geometry, salvador dali style, primary colors, detailed 8k"
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "ancient mystery creature, classical oil painting, many legs, dragon like, lightening bolts in the background, detailed"
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "it's van gogh's starry night but there is one giant flying creature in the sky spitting out fire, a screaming person in the foreground, artistic." I'm speechless that nano-banana actually cast van Gogh as the screaming person.
Left: SDXL model from Jan. 2024. Right: nano banana. Prompt: "Osamu Tezuka style, friendly monster, octopus like, big sparkly eyes, simple color palette." Maybe nano banana should read more comic books.
Like many people have already attested, nano-banana is versatile, fun, and easy to work with. Comparing with running a stable diffusion model in Colab just last year, which required a subscription upgrade, the Gemini image model runs smoothly on my 6-year-old laptop, and never had memory issue that I encountered frequently before. Pretty impressive stuff!
