13:06
Nano Banana
The video shows how the creator, inspired by a podcast demo from Google's Logan Kilpatrick, built a free custom product-photography app called 'Product Banana' inside Google AI Studio's code-assistant builder, using Nano Banana (Gemini's image model) as the generation backend — and it walks viewers through replicating and customizing their own version of that app.
The idea originated from a Greg Eisenberg podcast episode where Logan Kilpatrick (Google) demoed an app, built with AI Studio's code assistant, that offers 10 example environments and places a product into them using Nano Banana.
The creator built his own app, 'Product Banana,' by prompting Google AI Studio's code assistant over 'a couple of hours.'
The app workflow: upload a product image, choose lighting, choose aspect ratio, choose camera angle, then either write a custom prompt or give a short description that the app expands into a full prompt via a 'get ideas' button, then generate.
An edit-image function lets you request specific changes to an already-generated image (e.g., adding more raspberries around a can).
A prompt-refinement feature lets you type a short change (e.g., 'winter mountains') to regenerate a similar image in a new environment while keeping other selected settings.
A style-reference-image upload feature makes the app analyze a reference photo and integrate its elements (pose, environment, lighting) into the generated prompt.
Multi-parameter generation is imperfect: in one demonstrated attempt the model got the environment and camera angle right but added black borders instead of the correct aspect ratio; a follow-up regeneration fixed the aspect ratio but got the camera angle wrong.
Output images from the app have no watermark, unlike some other results the creator implies.
Building your own version starts at aistudio.google.com by clicking the 'Build' icon and describing the app to the code assistant; the video says viewers can pause and use an on-screen prompt to bootstrap a similar app.
The video recommends requesting only one or two features at a time from the code assistant so it doesn't get overwhelmed making too many changes at once.
Two features are called out as most important to add: (1) an explicit instruction that the app must reformat the uploaded image to the selected aspect ratio before sending it to Nano Banana, and (2) a history bar so previous generations aren't lost when a new one is made.
Without the aspect-ratio instruction, the video states output images come back in whatever aspect ratio the uploaded image was (square stays square, landscape stays landscape), regardless of the aspect ratio selected in the app.
There is a daily image-generation limit, described as 'quite high' — roughly 50 to 100 images per day — that resets daily.
The general building approach recommended is trial-and-error: use the app, notice what's missing or broken, and ask the code assistant to fix or add it, iterating toward something usable.
Nano Banana (Gemini image model) — The AI image-generation/editing model that powers the app's actual image creation from an assembled text prompt and uploaded product photo. Apply: Use it as the backend the app calls with the final assembled prompt to generate or edit the watermark-free product image.
Google AI Studio 'Build' (code assistant app builder) — A chatbot-driven app builder inside Google AI Studio (aistudio.google.com) that generates a working app from natural-language descriptions of desired features. Apply: Go to aistudio.google.com, click the Build icon, and describe the app and its features in chat to have the assistant generate and iterate on it.
One-or-two-features-per-prompt discipline — A build practice the video recommends of limiting each instruction to the code assistant to a single feature or at most two. Apply: When requesting new functionality from the code assistant, ask for one or two features at a time rather than a long combined list, to avoid the model mishandling too many simultaneous changes.
Aspect-ratio pre-formatting instruction — An explicit rule added to the app telling it to reformat the uploaded product image to match the user-selected output aspect ratio before it's sent to Nano Banana. Apply: Instruct the code assistant that the app must format the uploaded image according to the aspect ratio chosen in the UI before sending it to Nano Banana, since without this the output inherits the aspect ratio of the uploaded image instead.
History bar — A UI panel that preserves previously generated images instead of letting them disappear when a new one is generated. Apply: Ask the code assistant to add a history bar/section that stores and displays prior generations for comparison and reuse.
Style-reference-image analysis — A feature where an uploaded reference photo is analyzed for its visual elements and those elements are integrated in detail into the text prompt sent to Nano Banana. Apply: Instruct the code assistant to have the app look at an uploaded style reference image, extract its elements, and fold them into the generated prompt so the output mimics the reference's composition and lighting.
Prompt refinement by short text edits — A feature letting the user type a brief change (e.g., 'winter mountains') to regenerate a similar image in a different environment while keeping other chosen parameters. Apply: Type a short replacement phrase into the refine field and regenerate to swap the scene/environment while preserving the previously selected lighting, aspect ratio, and camera angle.
Edit-image function — A post-generation feature for requesting a specific, targeted modification to an already-generated image. Apply: After generating an image, press the edit image button and describe the specific change you want (e.g., add more raspberries around the can) to get a revised version.
'Get ideas' prompt generation from short description — A feature where a few keywords (e.g., 'sunny poolside') are expanded by the app into a full, detailed image-generation prompt incorporating the selected lighting, aspect ratio, and camera angle. Apply: Type a short scene description, click 'get ideas' to auto-generate a detailed prompt, copy it into the prompt box, and click generate image.
The core technical gotcha the video isolates is that Nano Banana defaults to preserving the input image's aspect ratio rather than the app's selected output aspect ratio — you have to explicitly instruct the app layer to reformat the image before the model call, which is a non-obvious prompt-engineering step for anyone building on top of the model.
Multi-parameter generations (lighting + aspect ratio + camera angle simultaneously) are shown to be unreliable in a specific, demonstrated way — the model tends to satisfy some parameters correctly while dropping others in the same generation, which the video frames as an expected trial-and-error cost rather than a rare failure.
The style-reference feature works by having the app first analyze the reference image into descriptive text elements and fold those into the prompt, rather than passing the reference image directly as visual conditioning alongside the product photo — meaning fidelity to the reference depends on how well the app's own analysis step captures it in words.
The video traces a concrete chain of propagation for a no-code AI workflow: a Google employee's podcast demo (Kilpatrick, via Eisenberg) is independently reverse-engineered and rebuilt by a creator within the same video cycle, then repackaged as a build-along tutorial for a wider audience — illustrating how fast AI Studio 'vibe-coded' app patterns spread through creator content.
The recommendation to limit each build request to one or two features at a time functions as an implicit admission that the code-assistant model degrades in reliability when given compound, multi-feature instructions in a single turn.
«In this video, I'm going to show you how you can create images like this for free using a secret NanoBanana powered image app.»
— 00:00
«I was watching videos on Nano Banana on YouTube when I came across this podcast interview from Greg Eisenberg.»
— 00:34
«So I started to play around with the code assistant in Google AI studio. And a couple of hours later I can now present to you product banana, my very own product image app.»
— 00:45
«But most importantly, as you can see, there is no watermark on this image.»
— 00:08
«So this is kind of a trial and error process. It doesn't get it right always but you can just generate new ones if something goes wrong.»
— 00:29
«I would go maybe one feature or max two features at a time so the model doesn't get caught up in doing too many changes at once.»
— 00:09
«It's important that you give the instruction that the app should format your uploaded image according to the aspect ratio selected in the app before sending it to Nanobanana.»
— 00:45
«I think you can generate something along the lines of 50 to 100 images per day and the token should reset every single day.»
— 00:10
Reception
Viewers express strong enthusiasm and genuine gratitude for the accessible tutorial and tool, with active usage and positive outcomes despite minor technical setup challenges.
The video is a practical build-along tutorial that converts a podcast demo into a replicable Google AI Studio workflow, concentrating its real teaching value in two specific implementation tips (the aspect-ratio pre-formatting instruction and the history bar) while candidly noting that multi-parameter image generation is unreliable and requires trial-and-error regeneration.
13:06