English · Русский
Read any manga. Offline. In one tap.
On-device manga and comics translation for Android, from Japanese, Chinese, Korean, Latin-script or Cyrillic-script pages into English, Brazilian Portuguese, Spanish, French, German or Russian. Tap once over any manga page, in any app, and read it in your language. By default nothing leaves your phone.
Important
Before you install
- Android 8.0 or newer on a 64-bit ARM phone. Tested on an 8 GB phone (Galaxy S23), where the default model peaks at about 2 GB of RAM. Phones with less RAM are untested; tell us how yours does.
- About 1.1 GB of models download on first launch, over Wi-Fi or mobile data. Keep about 1.5 GB free. After that, Yomu never uses the network.
- A page takes seconds, not milliseconds: about 15 s on a Galaxy S23 with the default model.
- Japanese reads best. Other sources use a PP-OCRv5 line reader for horizontal text; pick it in the OCR slot. Machine translation from a small on-device model makes mistakes.
Changes in this fork since it branched from artsaraiva/yomu at ba91c31.
0.3.6
- Translate from the notification shade. Settings → Appearance offers a second way to start a translation: from Yomu's notification instead of the floating button, so nothing sits over the page. The notification has Translate and Stop buttons, and the shade closes before the page is captured.
- ONNX Runtime 1.31.1 for detection and OCR, with the newer Arm KleidiAI kernels, including SME2.
- Build: Android Gradle Plugin 9.5, Gradle 9.8.2, Kotlin 2.4.21, Hilt 2.60.1, KSP 2.3.12, Compose BOM 2026.09.00, OkHttp 5 and current AndroidX libraries; compileSdk 37.
0.3.5
- Many-column narration boxes are read. manga-ocr squeezes every crop to 224 × 224 px, so a caption with a dozen vertical columns came back garbled and was skipped as unsure. A block of four or more vertical columns is now read two columns at a time, right to left. A 13-column box went from 0.74 confidence and garbage to 0.93 and clean text.
- Long one-line captions are read. A single line many times wider than tall, such as white text on a black bar, was squeezed flat the same way. It is now cut between characters into pieces about five characters long. A test caption went from 0.68 to 0.93.
0.3.1
- Choice of bubble detector. The Detection slot also offers
ogkalu/comic-text-and-bubble-detector(RT-DETR-v2 v4-s, int8, 11 MB, Apache-2.0), marked Experimental. It finds text in rectangular captions that the default YOLO26n misses, for example 0.95 against 0.09 on a boxed monologue. YOLO26n stays the default. See ADR-0020. - A new detector takes over at once. Picking another detector used to need an overlay restart; now the next page uses it.
0.2.35
- Russian target. Russian joins English, Brazilian Portuguese, Spanish, French and German.
- More source scripts. Chinese and English, Korean, Latin-script and Cyrillic-script pages are read with PP-OCRv5 line recognizers. The OCR model you pick sets the source language, and Japanese keeps manga-ocr.
- TranslateGemma 4B joins the translation catalog as Experimental and Multilingual, translating line by line in its own trained prompt format.
- Gemma 4 is labelled Multilingual, so it can be picked for every target language, Russian included.
- Build: Android Gradle Plugin 9.1, Gradle 9.3.1, built-in Kotlin, Hilt 2.59.2, KSP 2.3.6 and Room 2.8.5.
| Before | After one tap |
|---|---|
Unedited output from the default model (Qwen2.5 1.5B), recorded in real time on an Android emulator. The emulator runs on a desktop CPU, so a phone takes longer. Manga: ブラックジャックによろしく 佐藤秀峰 (Give My Regards to Black Jack, SHUHO SATO), used under the author's free secondary-use licence.
- What's new
- Features
- How it works
- Models
- Getting started
- Usage
- Project structure
- Testing
- Documentation
- Contributing
- License
- Works over any app. A floating button sits on top of your reader, browser or gallery. Yomu captures the screen only when you tap it.
- Six target languages. English, Brazilian Portuguese, Spanish, French, German or Russian. Switch from settings or the floating button.
- Five source families. Japanese (manga-ocr), Chinese and English, Korean, Latin script and Cyrillic script (PP-OCRv5). The OCR model you pick sets the source.
- On-device by default. Bubble detection, OCR and translation all run locally. No account, no API key, and no usage after the model download. An optional Network Model (Settings > Providers) can translate on your own Ollama, llama.cpp or OpenAI-compatible server instead; only the page text is sent (see ADR-0022).
- Page-level context. Every bubble on the page goes to the LLM in one call, in reading order, so pronouns and tone hold across a conversation.
- Manga-aware typesetting. Font size and line wrapping are fitted to each bubble's bounds.
- Choice of model. Pick from a curated list of GGUF translation models. Models that won't fit your device's RAM are hidden.
- Cancel any time. Tap the floating button again to stop a translation that's still running.
- Tunable. Bubble confidence threshold, font scale, generation profile, resource limits, and a live resource readout.
flowchart LR
A[Tap floating button] --> B[Screen capture<br/><sub>MediaProjection</sub>]
B --> C[Bubble detection<br/><sub>RT-DETR v4-s · ONNX</sub>]
C --> D[OCR per bubble<br/><sub>manga-ocr · ONNX</sub>]
D --> E[Context assembly<br/><sub>panels + reading order</sub>]
E --> F[Translation<br/><sub>GGUF LLM · llama.cpp</sub>]
F --> G[Typesetting]
G --> H[Overlay on the live page]
| Stage | Where | What it does |
|---|---|---|
| Capture | app/capture |
Grabs a single frame of the screen through MediaProjection |
| Detection | pipeline/bubble |
Finds text inside and outside speech bubbles with an RT-DETR detector, or with a single-class YOLO detector if chosen |
| OCR | pipeline/ocr |
Reads each cropped bubble: manga-ocr for Japanese, or a PP-OCRv5 recognizer line by line for other sources |
| Context | pipeline/context |
Groups bubbles into panels and orders them right-to-left, top-to-bottom |
| Translation | ml/ |
Sends the whole page to the LLM in the translation slot, using grammar-constrained, id-keyed output where the model supports it |
| Typesetting | pipeline/typesetting |
Wraps and sizes the English text to fit each bubble |
| Render | app/overlay |
Draws the typeset bubbles over the page in a system overlay window |
Yomu downloads every model on first launch. Each download is pinned to an exact Hugging Face revision.
| Slot | Model | Licence |
|---|---|---|
| Bubble detection (default) | ogkalu/comic-text-and-bubble-detector (RT-DETR-v2 v4-s, int8 ONNX); also finds text in rectangular captions |
Apache-2.0; see ADR-0020 |
| Bubble detection (alternative) | Kiuyha/Manga-Bubble-YOLO (yolo26n, ONNX) |
AGPL-3.0, from its Ultralytics YOLO26 base; see ADR-0007 |
| OCR | artsaraiva/manga-ocr-2025-onnx-past (manga-ocr-2025 with past key values) |
Apache-2.0 |
| OCR (other sources) | PaddlePaddle PP-OCRv5 mobile recognizers (Chinese and English, Korean, Latin, Cyrillic; ONNX) | Apache-2.0 |
| Translation (default) | Qwen2.5 1.5B Instruct Q4_K_M | Apache-2.0 |
| Translation | CAT-Translate 1.4B Q4_K_M | MIT |
| Translation (low storage) | CAT-Translate 0.8B Q4_K_M | MIT |
| Translation (custom) | Any GGUF you provide. Allowed but unsupported. | Its own |
Why these models were picked: ADR-0010, ADR-0016, ADR-0001.
Download yomu-<version>.apk from the latest release on your phone and open it. Android asks once to allow installs from your browser or file manager.
To get updates automatically, install it with Obtainium instead: tap the badge above, or add https://github.com/iBog/yomu as an app in Obtainium.
| Android device | Android 8.0 (API 26) or newer, arm64-v8a or x86_64 |
| JDK | 17 |
| Android SDK | compileSdk 36, plus NDK and CMake 3.22.1 (for llama.cpp) |
| Storage | about 1.5 GB free for the default models |
git clone --recursive https://github.com/iBog/yomu.git
cd yomu
# Already cloned without --recursive? Fetch llama.cpp:
git submodule update --init --recursive
# local.properties must point at your SDK
echo "sdk.dir=$HOME/Library/Android/sdk" > local.properties
./gradlew assembleDebug
./gradlew installDebug- Open Yomu and let setup download the detection, OCR and translation models.
- Grant permissions when asked: Display over other apps for the floating button, and Screen capture for reading the page.
- Start the overlay from the Home screen.
- Open a manga page in any app.
- Tap the floating button. Status appears while it runs, then translated bubbles cover the Japanese ones.
- Tap again to cancel. To stop the overlay, hold the button and tap Stop service.
Settings let you pick the target language, switch translation models, adjust the bubble confidence threshold and font scale, and cap the CPU and memory the model uses.
yomu/
├── app/ Android app: Compose UI, overlay service, screen capture, model downloads (Hilt, Room)
├── pipeline/ Page pipeline: detection → OCR → context → translation → typesetting
├── ml/ Inference: ONNX Runtime wrapper, llama.cpp JNI bridge (llama.cpp as a submodule)
├── core/ Shared types, constants, Result
├── scripts/ Device speed benchmark
└── docs/ ADRs, design system, agent docs
./gradlew lint test assembleDebug # what CI runs on every pull request
./gradlew connectedAndroidTest # instrumented tests (needs a device or emulator)
scripts/run-speed-benchmark.sh # per-stage timings and peak memory for every catalog modelThe speed benchmark only reports numbers and never fails on them. Run it on a real phone when you change a model or the inference runtime, since emulator timings don't match phones.
| Doc | Contents |
|---|---|
CONTEXT.md |
Domain language: deliverable, fit budget, translation slot, typeset bubble |
docs/adr/ |
Architecture decision records: why each model, threshold and trade-off was chosen |
docs/design/paper-mache-visual-system.md |
The paper-mâché visual system for the app's screens and overlay controls |
AGENTS.md |
Git, issue and code conventions |
Bug reports, ideas and pull requests are welcome. Ask questions in Discussions. CONTRIBUTING.md covers building, the issue-first workflow and what CI checks.
Yomu is licensed under the GNU Affero General Public License v3.0 (ADR-0019). Contributions are accepted under the same licence; there is no CLA. Each model is covered by its own licence (see Models).