D.01
Visual memory
A page request is multi-modal. It sends the text prompt, then the previous pages as images, then the sheets for the characters and places on that page, and tells the model to use them for continuity.
Hackathon entry · Google Nano Banana (Gemini 2.5 Flash Image) Hackathon · live on Vercel
NanoManga Studio
A studio that takes a one-line idea to finished, downloadable manga pages with Gemini.
01 // Problem
An image model draws one picture at a time and forgets the last one. A manga needs the same faces, clothes and places to hold from page to page, so each new page has to see what came before it.
- Gemini models
- 2
- API routes
- 7
02 // Architecture
- L1Planninggemini-2.5-flash acts as story editor and returns the whole plan as strict JSON: summary, characters, environments and a page-by-page plot.
- L2AssetsThe image model draws character sheets and empty environment shots from the plan, under separate concept-artist and background-artist prompts.
- L3Pagesgemini-2.5-flash-image-preview draws each page from a panel script plus earlier pages and the relevant asset sheets.
- L4Next.js route handlersSeven API routes behind the UI: one for image generation, six for the inspire and planning steps.
03 // Key_Decisions
D.02
The plan is the source of truth
Everything downstream reads the JSON plan. You can edit names, plot points and descriptions in it before a single image is drawn.
D.03
Pages stay editable
Click any page to view it full size, download it, or change its prompt and redraw it. The redraw still sees the pages before it.
D.04
Models are configuration
The text and image model names come from environment variables, so a newer model is a deploy setting and not a code change.
04 // From_The_Source
05 // Stack
- Next.js
- TypeScript
- Tailwind CSS
- Gemini API