Kumanga

Inspiration

Most AI art tools begin with a text prompt and replace the creator's composition. Kumanga begins with the creator instead: a rough Apple Pencil sketch, a story idea, and an intentional panel layout. I wanted people who cannot draw polished manga to experience the magic of finishing a page without giving up authorship.

What it does

Kumanga is a native iPad manga studio. A creator can sketch a character or scene, describe the intended design, add a visual reference, and transform the draft into polished manga art while preserving its composition. Generated characters become reusable assets, so the same hero can appear consistently across multiple panels. The result stays editable: creators can regenerate variations, change backgrounds or expressions, arrange panel layouts, add manga text elements, and export the finished page.

How I built it

Kumanga is written in Swift and SwiftUI, with PencilKit through a UIKit bridge for low-latency Apple Pencil drawing. The app has project, episode, character, and page models with local persistence; asynchronous AI generation is routed through AkarionHub, with multiple image-provider options including OpenAI GPT Image 2. Sign in with Apple, StoreKit subscriptions, credit tracking, sharing, and four interface languages are integrated into the same native workflow.

Codex was my primary engineering collaborator. I used Codex to inspect and modify the repository, design the architecture, implement PencilKit and persistence, integrate the AI service, debug complex SwiftUI state, localize the app, run command-line builds, and prepare TestFlight releases. GPT-5.6 powered those Codex sessions and helped reason across the large UI surface, service layer, persistence models, and release workflow. GPT-5.6 was the engineering model used through Codex; Kumanga's runtime image generation is a separate provider layer.

Challenges

The hardest problem was maintaining a coherent creative state across projects, episodes, reusable assets, AI jobs, and editable page elements. Async image generation also had to survive long requests without duplicate billing or lost results. The interface had to expose powerful controls without making a first-time creator feel like they were operating a professional desktop tool.

Accomplishments

Kumanga now supports an end-to-end path from a tiny sketch to a reusable character and a composed manga page. It is a real signed iPad app with StoreKit, authentication, persistence, localization, App Store delivery, and a focused creator-first AI workflow.

What I learned

The model is only one part of a creative product. The real value is the workflow around it: preserving intent, making results editable, and helping people iterate without losing ownership of their story.

What's next

Next I want to add more precise selection-based repainting, version history, richer dialogue tools, collaborative projects, and stronger cross-page character consistency.

Built With

• ios
• storekit
• swift
• swiftui
• testflight