Goes further with Koharu, the local-first manga translator with AI agents built in!
retained text
Hi LocalLLaMA, I have spent more than two months rebuilding Koharu to be the most powerful manga translator! TLDR: [https://github.com/koharu-rs/koharu](https://github.com/koharu-rs/koharu) (MIT or Apache 2.0, at your choice) It's been a while since I posted the Koharu update here. This time, I have added more features and built an established editor to make it feel more like Photoshop or Clip Studio Paint. You can type directly onto the page, restyle selected characters, and reshape text frames. There are also layers, history, marquee and lasso selection, Generative Fill, and a Remove brush. The canvas is GPU-rendered now, so it stays smooth even on big chapters. Everything still runs locally. LLMs run in-process through llama.cpp (Index-Translate 9B by default, plus Gemma 4, Qwen 3.x and LFM3), and inpainting uses LaMa, or FLUX.3 Klein / Qwen-Image through stable-diffusion.cpp. If you already run Ollama or LM Studio, you can point Koharu at that instead. On my machine, a page goes from raw to lettered in about 1 second once the models are loaded. I also added an agent you can ask to proofread or fix pages. It's exposed over MCP too, so you can drive Koharu from Claude Code or Codex if you'd rather. It works on CUDA, ROCm, Metal, and Vulkan. I'd like to hear your feedback or any form of suggestion!
document fetched 2026-11-19 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzhaao/goes_further_with_koharu_the_localfirst_manga