**Koharu** is an ML-powered manga translator written entirely in **Rust**, released under **Apache-2.0** by the developer **mayocream**. It has gathered **5,362 stars and 355 forks**, and development is relentless: releases **0.78.0 and 0.78.1** both landed on **22 August 2026**, with 0.77.5 the day before. The pitch is a local-first workflow that chains four separate model families — **object detection, OCR, inpainting, and LLM translation** — into one desktop application, with the explicit design note that vision models and LLMs run **locally** to keep source material private. The pipeline is the substance here, because manga translation is not a single ML problem. Detection and segmentation find text regions, speech bubbles, and cleanup masks using **Koharu Layout RF-DETR Seg 2XL**. OCR then reads the source text, with a choice of **PaddleOCR VL 1.6**, the project's own **Manga OCR**, or **Baberu OCR**. Inpainting reconstructs the artwork behind the removed lettering — the step that separates a usable page from an obvious patch job — and Koharu offers four options spanning generative and classical approaches: **FLUX.2 Klein**, **RORem mixed**, **LaMa**, and **AOT GAN**. Only then does translation run. Crucially, the pipeline is **selective**: any stage can be run at page or project scope, so a user can re-OCR one panel without redoing an entire volume. Translation is where the project's model-agnosticism shows. Local **GGUF** inference covers a wide catalog — LFM 2.5, Ministral 3, and multiple sizes of **Gemma 4** and **Qwen 3.5 / 3.6 / 3.8** — and hosted providers include Atlas Cloud, OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek, and OpenRouter. The README also lists uncensored community fine-tunes of Gemma 4 and Qwen as selectable local models, a pragmatic acknowledgement that a general-purpose assistant model will refuse to translate a meaningful share of published manga. Hardware coverage is broader than most Rust ML desktop apps attempt: **CUDA 13.0** (requiring an R580 or newer NVIDIA driver), **ROCm/HIP** for AMD on Windows and Linux, **Metal** on Apple silicon, and **Vulkan** as a cross-vendor alternative. When no accelerator initializes, it falls back to CPU with no GPU SDK required — slower, but it runs. The editor canvas is **WebGPU**-based inside an embedded CEF webview, which is worth flagging as a real requirement: WebGPU needs a current graphics driver *even when inference runs on the CPU*, so "CPU-only" does not mean "driver-free". Beyond the models, the application side is more finished than the version number suggests. It handles multi-format projects — raster images, archives, and PDFs with page sequencing — and does **multilingual text shaping and layout** with automatic fitting, font fallback, vertical CJK, and right-to-left support, which is the unglamorous work that most translation tools get wrong. Output can be a flattened delivery file or a **layered PSD** for further editing, and there is a proofreading pass for correcting OCR and translation output before export. Recent versions add an **agent-based workflow** for project inspection, editing, and pipeline control. The honest caveats are setup weight and version velocity. Assembling detection, OCR, inpainting, and an LLM means downloading several models before the first page is processed, and the per-backend runtime requirements vary by GPU vendor and operating system. At **0.78.x** with multiple releases landing per week, the project is still moving fast enough that pinning a version is wise for anyone mid-project.