shizuku-translator is a self-hosted pipeline for translating Japanese and Chinese manga or doujinshi into English — entirely on your own hardware. What makes this one stand out from the dozens of "ChatGPT-wrapper manga translators" is its target deployment: a Proxmox LXC container leveraging an AMD Radeon integrated GPU via llama.cpp's Vulkan backend. That's a refreshingly unusual stack in a world where everyone assumes CUDA.

The architecture pieces together several specialized components:

• comic-text-detector — finds speech bubbles and text regions on each page
• manga-ocr / PaddleOCR — extracts the original JP/CN text
• LaMa inpainting — cleanly erases the source text from the artwork
• llama.cpp (Vulkan) — runs a local LLM for the actual translation
• FastAPI — wraps it all into a service you can hit from anywhere on your LAN

This is genuinely interesting because it solves a real problem (the existing manga translation tools are either cloud-dependent, paid services, or require beefy NVIDIA GPUs) using a creative hardware path. Vulkan inference on AMD iGPUs is still a frontier area, and seeing a working end-to-end application built on it is valuable as a reference implementation.

Who would benefit:

• Homelabbers running Proxmox who want to put their AMD APUs to work
• Translators and scanlation groups who need a private, batch-capable pipeline
• Developers researching practical LLM inference on non-CUDA hardware
• Anyone curious about chaining specialized CV models (detection → OCR → inpainting → generation) into a coherent service

Even if you don't read manga, the LXC + Vulkan + llama.cpp deployment recipe is reusable for any local AI workload on AMD silicon.