[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-73299":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":20,"compositeScore":21,"rankGlobal":10,"rankLanguage":10,"license":22,"archived":23,"fork":23,"defaultBranch":24,"hasWiki":23,"hasPages":25,"topics":26,"createdAt":10,"pushedAt":10,"updatedAt":37,"readmeContent":38,"aiSummary":39,"trendingCount":16,"starSnapshotCount":16,"syncStatus":40,"lastSyncTime":41,"discoverSource":42},73299,"koharu","mayocream\u002Fkoharu","mayocream","ML-powered manga translator, written in Rust.","https:\u002F\u002Fkoharu.rs",null,"Rust",4623,261,14,90,0,11,80,284,60,29.25,"GNU General Public License v3.0",false,"main",true,[27,28,29,30,31,32,33,34,35,36],"computer-vision","deep-learning","gpu","gui","japanese","manga","rust","tauri","text-rendering","translation","2026-06-12 02:03:11","\u003Ch1 align=\"center\">Koharu\u003C\u002Fh1>\n\n\u003Cp align=\"center\">ML-powered manga translator, written in \u003Cb>Rust\u003C\u002Fb>.\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n\u003Cimg alt=\"GitHub Downloads (all assets, all releases)\" src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fdownloads\u002Fmayocream\u002Fkoharu\u002Ftotal?style=for-the-badge&link=https%3A%2F%2Fgithub.com%2Fmayocream%2Fkoharu%2Freleases%2Flatest\">\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n\u003Ca href=\"https:\u002F\u002Ftrendshift.io\u002Frepositories\u002F20649\" target=\"_blank\">\u003Cimg src=\"https:\u002F\u002Ftrendshift.io\u002Fapi\u002Fbadge\u002Frepositories\u002F20649\" alt=\"mayocream%2Fkoharu | Trendshift\" style=\"width: 250px; height: 55px;\" width=\"250\" height=\"55\"\u002F>\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n\u003Ca href=\"https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Finstall-koharu\u002F\" target=\"_blank\">Getting Started\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002F\" target=\"_blank\">Docs\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmayocream\u002Fkoharu\u002Fissues\" target=\"_blank\">Bug reports\u003C\u002Fa> · \u003Ca href=\"https:\u002F\u002Fdiscord.gg\u002FmHvHkxGnUY\" target=\"_blank\">Discord\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp align=\"center\">\n\u003Ca href=\"https:\u002F\u002Fkoharu.rs\u002Fja-JP\u002F\" target=\"_blank\">日本語\u003C\u002Fa> | \u003Ca href=\"https:\u002F\u002Fkoharu.rs\u002Fzh-CN\u002F\" target=\"_blank\">简体中文\u003C\u002Fa>\n\u003C\u002Fp>\n\nKoharu introduces a local-first workflow for manga translation, utilizing the power of ML to automate the process. It combines the capabilities of object detection, OCR, inpainting, and LLMs to create a seamless translation experience.\n\nUnder the hood, Koharu uses [candle](https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Fcandle) and [llama.cpp](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp) for high-performance inference, with [Tauri](https:\u002F\u002Fgithub.com\u002Ftauri-apps\u002Ftauri) for the desktop app. All components are written in Rust, ensuring safety and speed.\n\n> [!NOTE]\n> Koharu runs its vision models and LLMs **locally** on your machine to keep your data private and secure.\n\n---\n\n![screenshot](docs\u002Fen-US\u002Fassets\u002Fkoharu-screenshot-en.png)\n\n> [!NOTE]\n> Support and discussion are available on the [Discord server](https:\u002F\u002Fdiscord.gg\u002FmHvHkxGnUY).\n\n## Features\n\n- Automatic detection of text regions, speech bubbles, and cleanup masks\n- OCR for manga dialogue, captions, and other page text\n- Inpainting to remove source lettering from the page\n- Translation with local or remote LLM backends\n- Advanced text rendering with vertical CJK and RTL support\n- Codex image-to-image generation for end-to-end page redraws from a source image and prompt\n- Layered PSD export with editable text\n- Local HTTP API and MCP server for automation\n\nFor installation and first-run guidance, see [Install Koharu](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Finstall-koharu\u002F) and [Translate Your First Page](https:\u002F\u002Fkoharu.rs\u002Ftutorials\u002Ftranslate-your-first-page\u002F).\n\n## Usage\n\n### Hotkeys\n\nCanvas:\n\n- \u003Ckbd>Ctrl\u003C\u002Fkbd> + Mouse Wheel: Zoom in\u002Fout\n- \u003Ckbd>Ctrl\u003C\u002Fkbd> + Drag: Pan the canvas\n\nTools:\n\n- \u003Ckbd>V\u003C\u002Fkbd>: Select tool\n- \u003Ckbd>M\u003C\u002Fkbd>: Block tool\n- \u003Ckbd>B\u003C\u002Fkbd>: Brush tool\n- \u003Ckbd>E\u003C\u002Fkbd>: Eraser tool\n- \u003Ckbd>R\u003C\u002Fkbd>: Repair Brush tool\n- \u003Ckbd>[\u003C\u002Fkbd> \u002F \u003Ckbd>]\u003C\u002Fkbd>: Decrease \u002F increase brush size\n\nHistory and selection:\n\n- \u003Ckbd>Ctrl\u003C\u002Fkbd> + \u003Ckbd>Z\u003C\u002Fkbd> \u002F \u003Ckbd>Cmd\u003C\u002Fkbd> + \u003Ckbd>Z\u003C\u002Fkbd>: Undo\n- \u003Ckbd>Ctrl\u003C\u002Fkbd> + \u003Ckbd>Shift\u003C\u002Fkbd> + \u003Ckbd>Z\u003C\u002Fkbd> \u002F \u003Ckbd>Cmd\u003C\u002Fkbd> + \u003Ckbd>Shift\u003C\u002Fkbd> + \u003Ckbd>Z\u003C\u002Fkbd>: Redo\n- \u003Ckbd>Ctrl\u003C\u002Fkbd> + \u003Ckbd>A\u003C\u002Fkbd> \u002F \u003Ckbd>Cmd\u003C\u002Fkbd> + \u003Ckbd>A\u003C\u002Fkbd>: Select all text blocks on the current page\n\nFor the full list and customization details, see [Keyboard Shortcuts](https:\u002F\u002Fkoharu.rs\u002Freference\u002Fkeyboard-shortcuts\u002F).\n\n### Export\n\nKoharu can export the current page either as a flattened rendered image or as a layered Photoshop PSD. PSD export preserves helper layers and writes translated text as editable text layers, which is useful for downstream cleanup and manual refinement.\n\nFor export behavior, PSD contents, and file naming, see [Export Pages and Manage Projects](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Fexport-and-manage-projects\u002F).\n\n### MCP Server\n\nKoharu includes a built-in MCP server for local agent integrations. By default it listens on a random local port, but you can pin it with `--port`.\n\n```bash\n# macOS \u002F Linux\nkoharu --port 9999\n# Windows\nkoharu.exe --port 9999\n```\n\nThen point your client at `http:\u002F\u002Flocalhost:9999\u002Fmcp`.\n\nFor local setup and the available tools, see [Run GUI, Headless, and MCP Modes](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Frun-gui-headless-and-mcp\u002F), [Configure MCP Clients](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Fconfigure-mcp-clients\u002F), and [MCP Tools Reference](https:\u002F\u002Fkoharu.rs\u002Freference\u002Fmcp-tools\u002F).\n\n### Headless Mode\n\nKoharu can run without launching the desktop window.\n\n```bash\n# macOS \u002F Linux\nkoharu --port 4000 --headless\n# Windows\nkoharu.exe --port 4000 --headless\n```\n\nYou can then connect to the web client at `http:\u002F\u002Flocalhost:4000`.\n\nFor runtime modes, ports, and local endpoints, see [Run GUI, Headless, and MCP Modes](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Frun-gui-headless-and-mcp\u002F).\n\n### Runtime Configuration\n\nKoharu lets you configure the shared local data path plus HTTP connect timeout, read timeout, and retry count used by downloads and provider requests.\n\nThose values are loaded at startup, so changing them saves the config and restarts the app.\n\n### Google Fonts\n\nKoharu includes built-in Google Fonts support for translated text rendering, so you can use web fonts without managing font files by hand.\n\nGoogle Fonts are fetched on demand from a bundled catalog. Koharu caches downloaded files under the app data directory and reuses them for later renders, so you usually only need an internet connection the first time a family is used on that machine. Once cached, a Google Font behaves like any other local render font.\n\n### Text Rendering\n\nKoharu includes a dedicated text renderer tuned for manga lettering, using Unicode-aware [OpenType](https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Ftypography\u002Fopentype\u002Fspec\u002F) shaping, script-aware line breaking, precise glyph metrics, and real glyph bounds instead of generic browser or OS text primitives.\n\nIt supports vertical CJK layout, right-to-left scripts, font fallback, vertical punctuation alignment, constrained-box fitting, and manga-oriented stroke and effect compositing so translated text reads naturally inside speech bubbles, captions, and other irregular page layouts.\n\n## GPU Acceleration\n\nKoharu supports CUDA, experimental ZLUDA, Metal, and Vulkan. CPU fallback is always available when the accelerated path is unavailable or not worth the setup cost on your system.\n\n### CUDA (NVIDIA GPUs on Windows and Linux)\n\nOn Windows and Linux, Koharu ships with CUDA support so it can use NVIDIA GPUs for the full local pipeline.\n\nKoharu bundles CUDA Toolkit 13.0. The required DLLs are extracted to the application data directory on first run.\n\n> [!NOTE]\n> Make sure you have current NVIDIA drivers installed. You can update them through [NVIDIA App](https:\u002F\u002Fwww.nvidia.com\u002Fen-us\u002Fsoftware\u002Fnvidia-app\u002F).\n\n#### Supported NVIDIA GPUs\n\nKoharu supports NVIDIA GPUs with compute capability 8.0 or higher.\n\nFor GPU compatibility references, see [CUDA GPU Compute Capability](https:\u002F\u002Fdeveloper.nvidia.com\u002Fcuda-gpus).\n\n### ZLUDA (AMD GPUs on Windows, experimental)\n\nKoharu supports experimental ZLUDA acceleration on Windows for AMD GPUs.\nZLUDA is a CUDA compatibility layer that lets some CUDA workloads run on AMD GPUs.\n\nTo use it, install the [AMD HIP SDK](https:\u002F\u002Fwww.amd.com\u002Fen\u002Fdeveloper\u002Fresources\u002Frocm-hub\u002Fhip-sdk.html).\n\n### Metal (Apple Silicon on macOS)\n\nKoharu supports Metal on Apple Silicon Macs. No extra runtime setup is required beyond a normal app install.\n\n### Vulkan (Windows and Linux)\n\nKoharu also supports Vulkan on Windows and Linux. This backend is currently used primarily for OCR and local LLM inference.\n\nDetection and inpainting still depend on CUDA, ZLUDA, or Metal, so Vulkan is useful but not a full replacement for the main accelerated path. AMD and Intel GPUs can still benefit from it.\n\n### CPU Fallback\n\nYou can always force Koharu to use CPU for inference:\n\n```bash\n# macOS \u002F Linux\nkoharu --cpu\n# Windows\nkoharu.exe --cpu\n```\n\nFor backend selection, fallback behavior, and model runtime support, see [Acceleration and Runtime](https:\u002F\u002Fkoharu.rs\u002Fexplanation\u002Facceleration-and-runtime\u002F).\n\n## ML Models\n\nKoharu uses a staged stack of vision and language models instead of trying to solve the entire page with a single network.\n\n### Computer Vision Models\n\nKoharu uses multiple pretrained models, each tuned for a specific part of the page pipeline.\n\n#### Detection and Layout\n\nThese models find text regions, speech bubbles, and page structure.\n\n- [anime-text-yolo](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Fanime-text-yolo) for text block detection\n- [comic-text-bubble-detector](https:\u002F\u002Fhuggingface.co\u002Fogkalu\u002Fcomic-text-and-bubble-detector) for joint text block and speech bubble detection\n- [comic-text-detector](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Fcomic-text-detector) for text segmentation masks\n- [PP-DocLayoutV3](https:\u002F\u002Fhuggingface.co\u002FPaddlePaddle\u002FPP-DocLayoutV3_safetensors) for document layout analysis\n- [speech-bubble-segmentation](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Fspeech-bubble-segmentation) for dedicated speech bubble detection\n\n#### OCR\n\nThese models recognize source text after detection.\n\n- [PaddleOCR-VL-1.5](https:\u002F\u002Fhuggingface.co\u002FPaddlePaddle\u002FPaddleOCR-VL-1.5) for OCR text recognition\n- [Manga OCR](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Fmanga-ocr) for OCR\n- [MIT 48px OCR](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Fmit48px-ocr) for OCR\n\n#### Inpainting\n\nThese models remove source lettering before translated text is rendered back onto the page.\n\n- [FLUX.2 Klein 4B](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FFLUX.2-klein-4B-GGUF) for FLUX.2-based inpainting\n- [lama-manga](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Flama-manga) for inpainting\n- [aot-inpainting](https:\u002F\u002Fhuggingface.co\u002Fmayocream\u002Faot-inpainting) for inpainting\n\n#### Font Analysis\n\nThis model helps infer source font and color characteristics for rendering.\n\n- [YuzuMarker.FontDetection](https:\u002F\u002Fhuggingface.co\u002Ffffonion\u002Fyuzumarker-font-detection) for font and color detection\n\nThe required models are downloaded automatically on first use.\n\nSome models are consumed directly from upstream Hugging Face repos, while Rust-friendly safetensors conversions are hosted on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fmayocream) when Koharu needs a converted bundle.\n\nFor a closer look at the pipeline, see [Models and Providers](https:\u002F\u002Fkoharu.rs\u002Fexplanation\u002Fmodels-and-providers\u002F) and the [Technical Deep Dive](https:\u002F\u002Fkoharu.rs\u002Fexplanation\u002Ftechnical-deep-dive\u002F).\n\n### Large Language Models\n\nKoharu supports both local and remote LLM backends. Local models run through [llama.cpp](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp) and are downloaded on demand. Hosted and self-hosted APIs are also supported when you want to use a provider instead of a downloaded model. When possible, Koharu also tries to preselect sensible defaults based on your system locale.\n\n#### General-Purpose Local Models\n\nThese are broad instruct models that work well when you want one local model for many translation tasks.\n\n- Gemma 4 instruct: [gemma4-e2b-it](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002Fgemma-4-E2B-it-GGUF), [gemma4-e4b-it](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002Fgemma-4-E4B-it-GGUF), [gemma4-26b-a4b-it](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002Fgemma-4-26B-A4B-it-GGUF), [gemma4-31b-it](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002Fgemma-4-31B-it-GGUF)\n- Qwen 3.5: [qwen3.5-0.8b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-0.8B-GGUF), [qwen3.5-2b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-2B-GGUF), [qwen3.5-4b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-4B-GGUF), [qwen3.5-9b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-9B-GGUF), [qwen3.5-27b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-27B-GGUF), [qwen3.5-35b-a3b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.5-35B-A3B-GGUF)\n- Qwen 3.6: [qwen3.6-27b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.6-27B-GGUF), [qwen3.6-35b-a3b](https:\u002F\u002Fhuggingface.co\u002Funsloth\u002FQwen3.6-35B-A3B-GGUF)\n\n#### NSFW-Capable Local Models\n\nThese variants relax the safety tuning applied to the corresponding base instruct models.\n\n- Gemma 4 uncensored: [gemma4-e2b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FGemma-4-E2B-Uncensored-HauhauCS-Aggressive), [gemma4-e4b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FGemma-4-E4B-Uncensored-HauhauCS-Aggressive)\n- Qwen 3.5 uncensored: [qwen3.5-2b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.5-2B-Uncensored-HauhauCS-Aggressive), [qwen3.5-4b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.5-4B-Uncensored-HauhauCS-Aggressive), [qwen3.5-9b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.5-9B-Uncensored-HauhauCS-Aggressive), [qwen3.5-27b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.5-27B-Uncensored-HauhauCS-Aggressive), [qwen3.5-35b-a3b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive)\n- Qwen 3.6 uncensored: [qwen3.6-27b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.6-27B-Uncensored-HauhauCS-Balanced), [qwen3.6-35b-a3b-uncensored](https:\u002F\u002Fhuggingface.co\u002FHauhauCS\u002FQwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive)\n\n#### Fine-Tuned Translation Models\n\nThese models are more specialized for translation quality, language coverage, or lower-resource setups.\n\n- [vntl-llama3-8b-v2](https:\u002F\u002Fhuggingface.co\u002Flmg-anon\u002Fvntl-llama3-8b-v2-gguf): a Q5_K_M GGUF, best when translation quality matters more than speed or memory use\n- [lfm2.5-1.2b-instruct](https:\u002F\u002Fhuggingface.co\u002FLiquidAI\u002FLFM2.5-1.2B-Instruct-GGUF): a smaller multilingual instruct model that is easier to run on CPUs or low-memory GPUs\n- [sugoi-14b-ultra](https:\u002F\u002Fhuggingface.co\u002Fsugoitoolkit\u002FSugoi-14B-Ultra-GGUF) and [sugoi-32b-ultra](https:\u002F\u002Fhuggingface.co\u002Fsugoitoolkit\u002FSugoi-32B-Ultra-GGUF): larger translation-oriented options when you have more VRAM or RAM available\n- [sakura-galtransl-7b-v3.7](https:\u002F\u002Fhuggingface.co\u002FSakuraLLM\u002FSakura-GalTransl-7B-v3.7): a smaller IQ4_XS GGUF, a good balance of quality and speed on 8 GB GPUs\n- [sakura-1.5b-qwen2.5-v1.0](https:\u002F\u002Fhuggingface.co\u002Fshing3232\u002FSakura-1.5B-Qwen2.5-v1.0-GGUF-IMX): lighter and faster, useful on mid-range GPUs or CPU-only setups\n- [hunyuan-mt-7b](https:\u002F\u002Fhuggingface.co\u002FMungert\u002FHunyuan-MT-7B-GGUF): a Q4_K_M GGUF with broad multilingual translation coverage\n\nLLMs are downloaded on demand when you activate a model. For constrained memory environments, start with a smaller model. When VRAM or RAM permits, 7B and 8B class models generally provide better translation quality.\n\n#### Cloud Providers\n\nKoharu supports hosted APIs from [OpenAI](https:\u002F\u002Fplatform.openai.com\u002F), [Gemini](https:\u002F\u002Fai.google.dev\u002F), [Claude](https:\u002F\u002Fwww.anthropic.com\u002Fapi), and [DeepSeek](https:\u002F\u002Fplatform.deepseek.com\u002F) instead of a local GGUF model.\n\nBuilt-in cloud catalogs include current text-output models for OpenAI, Gemini, Claude, and DeepSeek, including GPT-5.5\u002F5.4\u002F5.x, Gemini 3.1\u002F3\u002F2.5\u002F2.0, Claude Opus\u002FSonnet\u002FHaiku 4.x, DeepSeek V4, and compatibility aliases such as `deepseek-chat` and `deepseek-reasoner`.\n\n#### Codex Image-to-Image Generation\n\nKoharu can use Codex for end-to-end image-to-image generation. This workflow sends the current source page image plus a user prompt to Codex, then stores the generated image as a rendered page result.\n\nThis feature requires a ChatGPT account with Codex access. Two-factor authentication must be enabled on the account before device-code login can complete successfully.\n\nCodex image generation is useful when you want the model to translate visible text, remove the original lettering, and redraw the page in one pass. Because the image request is processed by the ChatGPT Codex backend, failures can include upstream OpenAI request IDs and may need to be retried.\n\n#### Machine Translation Providers\n\nFor pure machine-translation use cases, Koharu also supports [DeepL](https:\u002F\u002Fwww.deepl.com\u002F), [Google Cloud Translation](https:\u002F\u002Fcloud.google.com\u002Ftranslate), and [Caiyun](https:\u002F\u002Ffanyi.caiyunapp.com\u002F). These providers translate without an LLM-style chat or system prompt; you provide an API key and Koharu uses the upstream translate endpoint directly.\n\n#### OpenAI-Compatible Providers\n\nKoharu supports OpenAI-compatible endpoints such as LM Studio, OpenRouter, and other self-hosted or third-party APIs that expose `\u002Fv1\u002Fmodels` and `\u002Fv1\u002Fchat\u002Fcompletions`.\n\nCloud providers can be configured with API keys. OpenAI-compatible providers also need a custom base URL. API keys are stored securely in your system keychain instead of plain text config files. API keys are optional for local servers such as LM Studio, but are usually required for hosted services such as OpenRouter.\n\nUse a remote provider to avoid local model downloads, reduce VRAM or RAM requirements, or integrate with an existing hosted or self-hosted endpoint. Keep in mind that the OCR text selected for translation is sent to the provider you configured.\n\nFor LM Studio, OpenRouter, and other OpenAI-style endpoints, see [Use OpenAI-Compatible APIs](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Fuse-openai-compatible-api\u002F). For provider configuration, see [Settings Reference](https:\u002F\u002Fkoharu.rs\u002Freference\u002Fsettings\u002F).\n\n## Installation\n\nYou can download the latest release of Koharu from the [releases page](https:\u002F\u002Fgithub.com\u002Fmayocream\u002Fkoharu\u002Freleases\u002Flatest).\n\nWe provide prebuilt binaries for Windows, macOS, and Linux. For the standard install flow, see [Install Koharu](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Finstall-koharu\u002F). If something goes wrong, see [Troubleshooting](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Ftroubleshooting\u002F).\n\nKoharu can run offline with local models once the required runtimes, models, and fonts are already present on disk.\n\n### WinGet\n\nOn Windows, you can install Koharu with [winget](https:\u002F\u002Flearn.microsoft.com\u002Fen-us\u002Fwindows\u002Fpackage-manager\u002Fwinget\u002F):\n\n```bash\nwinget install koharu\n```\n\n### Docker\n\nKoharu also publishes official Docker images for headless use. You can pull the latest image from GitHub Container Registry:\n\n```bash\ndocker pull ghcr.io\u002Fmayocream\u002Fkoharu:latest\n```\n\nThen run the container with the desired port mapping:\n\n```bash\ndocker run -p 4000:4000 --gpus all ghcr.io\u002Fmayocream\u002Fkoharu:latest\n```\n\n## Troubleshooting\n\nKoharu provides a diagnostic mode that outputs detailed logs and system information to help identify issues with installation, GPU acceleration, model loading, and more. To enable it, run:\n\n```bash\n# macOS \u002F Linux\nkoharu --debug\n# Windows\nkoharu.exe --debug\n```\n\n## Development\n\nTo build Koharu from source, follow the steps below.\n\n### Prerequisites\n\n- [Rust](https:\u002F\u002Fwww.rust-lang.org\u002Ftools\u002Finstall) 1.95 or later (Rust 2024 edition)\n- [Bun](https:\u002F\u002Fbun.sh\u002F) 1.0 or later\n\nOptional dependencies for GPU acceleration builds:\n\n- [LLVM](https:\u002F\u002Fllvm.org\u002F) 15 or later (for GPU acceleration builds)\n- [CUDA Toolkit](https:\u002F\u002Fdeveloper.nvidia.com\u002Fcuda-13-0-0-download-archive) 13.0 (for CUDA and ZLUDA support on Windows)\n- [AMD HIP SDK](https:\u002F\u002Fwww.amd.com\u002Fen\u002Fdeveloper\u002Fresources\u002Frocm-hub\u002Fhip-sdk.html) (for ZLUDA support on Windows)\n\n### Install dependencies\n\n```bash\nbun install\n```\n\n### Development\n\n```bash\nbun dev\n```\n\n### Build\n\n```bash\nbun run build\n```\n\nThe built binaries are written to `target\u002Frelease`.\n\nFor platform-specific build notes, see [Build From Source](https:\u002F\u002Fkoharu.rs\u002Fhow-to\u002Fbuild-from-source\u002F). For the local development workflow, see [Contributing](https:\u002F\u002Fkoharu.rs\u002Fcontribute\u002Fintroduction\u002F).\n\n## Sponsorship\n\nIf Koharu is useful in your workflow, consider sponsoring the project.\n\n- [GitHub Sponsors](https:\u002F\u002Fgithub.com\u002Fsponsors\u002Fmayocream)\n- [Patreon](https:\u002F\u002Fwww.patreon.com\u002Fmayocream)\n\n## Contributors ❤️\n\nThanks to all the contributors who have helped make Koharu better!\n\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmayocream\u002Fkoharu\u002Fgraphs\u002Fcontributors\">\n  \u003Cimg src=\"https:\u002F\u002Fcontrib.rocks\u002Fimage?repo=mayocream\u002Fkoharu\" \u002F>\n\u003C\u002Fa>\n\n## License\n\nKoharu is licensed under the [GNU General Public License v3.0](LICENSE).\n","Koharu 是一个基于机器学习的漫画翻译工具，使用 Rust 语言编写。该项目集成了对象检测、OCR（光学字符识别）、图像修复以及大语言模型等技术，以实现自动化的文本区域检测、对话框识别、源文字清除及翻译等功能。Koharu 支持本地运行视觉模型和语言模型，确保用户数据的安全与隐私。此外，它还提供了高级的文字渲染功能，支持垂直CJK文字显示，并能够导出分层PSD文件以便进一步编辑。适用于需要高效且安全地进行多语言漫画翻译的场景。",2,"2026-06-11 03:44:54","high_star"]