AI9
Manga Colorizer

Demo

AI9 is the manga colorizer

Real local inference. Black-and-white manga in, colorized page out. This GIF is a real capture of AI9 taking a B&W page through the local GPU colorization workflow.

The Real Thing

Not a mockup

The demo above, the extension popup, and the live backend check were captured from this project's own repository and its actually-running install, not staged or drawn.

● Live backend check, this machine, just now

$ curl -sk https://127.0.0.1:5000/ Manga Colorizer is Up and Running!

This is the same install this page tells you how to run: a real Flask backend, a real GPU behind it, and the real Windows background task (MangaColorizerAI9) keeping it alive between crashes and reboots.

In the Otaconskeep Story

The GPU/compute side of the Keep

Otacon is the agent platform. Genome Voice Trainer clones Piper voices for those agents on your GPU. AI9 is the manga colorizer — proof of the same discipline on a standalone GPU pipeline: detect real hardware, select the exact runtime, fail loudly when something's wrong, never a silent cloud fallback. See the demo GIF above for B&W → color on a real local install.

Architecture

Extension → backend → GPU pipeline

Concurrency is 1 by default: one page colorizes at a time, queued, so the GPU stays available for other work. A two-tier cache keyed by image content hash and processing options means switching color presets on an already-read page reuses the raw GPU output instead of a full re-run.

Hardware Support

Automatic GPU detection & runtime selection

The installer reads your GPU name and driver's max CUDA version directly from nvidia-smi, then picks the matching PyTorch wheel. No manual CUDA version guessing.

Verified with a real GPU inference pass (matmul + cuDNN convolution kernels), not just torch.cuda.is_available(). That call can return True while still lacking compiled kernels for your specific architecture.

Install

Windows 10/11: double-click, not curl

AI9's installer targets Git Bash on Windows and refuses to run under WSL. It is not a Linux-style curl | bash one-liner. Here's the real supported path.

No WSL, no Linux VM, nothing to install ahead of time. If Git Bash isn't already on your machine, install_ai9.bat installs it for you automatically.

1. Download the repo

   GitHub's green Code button → Download ZIP (then extract it), or git clone it.

2. Double-click install_ai9.bat

   install_ai9.bat Copy

   It finds or installs Git Bash automatically, then runs the real installer, install_ai9.sh, inside it.

3. Or, if you already use Git Bash

   Open Git Bash in the extracted folder and run:

   ./install_ai9.sh Copy

4. Wait for it to finish

   Installs Python 3.12, Git, and Firefox if missing (via winget); detects your GPU and installs the matching PyTorch build; downloads and verifies the AI model weights; runs a real GPU inference test; registers a Windows background task so the server auto-starts and restarts itself on crash.

5. Load the Firefox extension (the one manual step)

   Firefox requires a human click for unsigned add-ons. No installer can do this part. Go to about:debugging#/runtime/this-firefox → Load Temporary Add-on → select extension/manifest.json from the install folder. This needs repeating after every full Firefox restart.

Full install documentation on GitHub →

Status

What ships, and its known limits

Known limitations (from the repository's own troubleshooting notes):

• The Firefox extension is loaded as an unsigned temporary add-on. It does not persist across a full Firefox restart and needs reloading.
• A canvas-based reader can be repainted with zero DOM mutation trace, so those pages use a narrow, low-frequency fingerprint poll instead of pure event-driven detection: a deliberate, disclosed tradeoff.
• The backend serves over a self-signed certificate; browsers show a one-time trust warning per origin.