Manga Colorization and Translation with Character & Background Consistency
Nano Manana is an automated system for manga colorization and translation that maintains consistency of characters and backgrounds across entire pages. Built on Gemini Imagen 3, it addresses the key challenges in manga processing:
- Colorization: Transforms black-and-white manga into vibrant, professionally colored pages
- Translation: Translates text including speech balloons, onomatopoeia, and background text directly into the image
- Consistency: Maintains character appearance (eye color, hair color, skin tone, clothing) across multiple pages
Achieves up to 10x speedup for translation and 5x speedup for colorization compared to sequential processing.
Professional-quality coloring with depth, shading, and highlights through optimized prompt engineering.
Supports displaying both original and translated text simultaneously for language learning or translation verification.
Easy-to-use web interface deployed at nano-manana.vercel.app
| Feature | MangaDiT | MangaNinja | Nano Manana |
|---|---|---|---|
| Reference Required | ✅ Yes | ✅ Yes | ❌ No |
| Multi-panel Support | ❌ No | ❌ No | ✅ Yes |
| Translation | ❌ No | ❌ No | ✅ Yes |
| Low VRAM Support | ❌ No | ✅ Yes (API) | |
| Batch Processing | ❌ No | ❌ No | ✅ Yes |
Input Pages → Batch Processing → Parallel API Calls → Translated Pages
Input Pages → Reference Selection → Batch API Calls (with prev. results) → Colorized Pages
The colorization pipeline uses previously completed pages as references to maintain consistency:
min(batchNumber, batchSize, completedCount)references per batch- Maximum 5 references due to API limitations
While Nano Manana achieves high-quality results in most cases, some limitations exist due to the underlying Gemini Imagen 3 model:
- Style Drift: Occasionally generates completely different art styles
- Scene Generation: May create scenes not present in the original
- Aspect Ratio: Sometimes outputs different aspect ratios despite explicit instructions
These issues can be mitigated through a user-confirmation workflow with automatic regeneration.
| Metric | Average |
|---|---|
| Total Tokens | 2,542.30 |
| Total Cost | $0.14 |
- Backend: Gemini Imagen 3 API
- Frontend: Web application deployed on Vercel
- MangaDiT - Reference-Guided Line Art Colorization with Hierarchical Attention
- MangaNinja - Line Art Colorization with Precise Reference Following
- Context-Informed Manga Translation - Machine Translation using Multimodal LLMs
This project is for educational and research purposes.
This project was developed as part of a Deep Learning course at Yonsei University (2025-2).
Try it now: nano-manana.vercel.app