I built a manga colouriser for a friend, and it runs fully on my laptop

Parth Mittal — Sun, 04 Oct 2026 16:47:22 +0000

My friend Vinay reads a lot of manga, and it's all black and white. For the Hacktoberfest "Build for a Friend" challenge I built a tool that colours manga pages and keeps each character looking the same from page to page.

What it does

You upload black-and-white pages (images, PDF or CBZ) and add each main character once, with a few coloured reference pictures. The app then:

Finds the panels and speech bubbles on every page.
Finds the characters in each panel and works out who they are.
Colours each panel, using the character's reference picture as a guide.
Reads the character's hair, skin and clothes colours from the references and nudges every page back to those colours, so a skirt doesn't change from red to blue between pages.

Speech bubbles and lettering are left untouched. Any wrong character guess can be fixed by hand with a dropdown.

![The app's interface running in Quick preview mode on CPU
The interface, running in Quick preview mode (no AI) on a CPU.
u can visit the interface from here :https://manga-colorizer-ahed6hjbysmjku3vfzjop9.streamlit.app

Note: the hosted version is only for looking around. You can try the interface and see the panel and speech-bubble mark-ups on a page. It does not include the AI models, so it can't do the full colouring. For that, clone the repo from GitHub and run it on your own computer, which downloads the models (about 6 GB, one time). Anything you upload to the hosted demo goes to Streamlit's servers, so only use pages you're happy to share.

Why open-source AI matters here

It can run on your own computer. The full version is built to run locally, so someone's unpublished pages never have to be uploaded to a company's server.
It costs nothing to run. No per-page fees or subscriptions.
I can swap models. Stable Diffusion with a line-art ControlNet and IP-Adapter does the colouring, OWL-ViT finds characters, CLIP names them, and SegFormer finds hair and clothes. If one part is weak, I can replace only that part.

How the colours stay consistent

Each character has a "memory": reference pictures plus a stored colour for each body part (hair, skin, top, skirt, shoes and so on). The reference steers the AI, the colours are turned into words in the prompt, and after colouring a clothes-segmentation model shifts each part toward its stored colour.

What I got working, and what I didn't

Working and tested: panel detection, speech-bubble protection, the character memory, the colour-correction step, the Streamlit interface, and the failure handling. I tested the pipeline with stand-in models, including the case where one panel fails and the rest of the page is still coloured.
Not done: I wasn't able to run the full AI colouring end-to-end before the deadline, because the models are about 6 GB. So I can't show AI-coloured results here. The Quick preview mode in the screenshot only fills regions with flat colours and is meant for checking the layout.
Weakest part: working out which character is in which panel, which is why you can correct it by hand.

Code:https://github.com/afk-Parth/manga-colorizer

DEV Community: Parth Mittal

I built a manga colouriser for a friend, and it runs fully on my laptop

What it does

Why open-source AI matters here

How the colours stay consistent

What I got working, and what I didn't