Yomikae: translate raw manga in your browser, lettered into the bubbles like a real scanlation
retained text
I read a lot of untranslated manga/doujinshi on pixiv and X (you can also create an adapter for aggregator sites like tachiyomi), so I built Yomikae. You browse normally, and each page gets the Japanese erased and English lettered into the speech bubbles, like a scanlation. It runs on your Claude subscription through `claude -p`, with no API key. Only the text goes to Claude, and the page images stay on your computer. A local LLM or a codex sub works too. GitHub (GPL-3.0): [https://github.com/gravewhisper/yomikae](https://github.com/gravewhisper/yomikae) Benchmarks: |Benchmark|Yomikae|Koharu|manga-image-translator|BallonsTranslator| |:-|:-|:-|:-|:-| |Names spelled right (of 113)|**86**|70|45|65| |OCR error rate|**1.0%**|1.3%|4.3%|5.6%| |Translation quality (COMET)|0.747|**0.752**|0.731|0.736| |Lines translated (of 1586)¹|1546|**1583**|1498|1530| |Seconds per page (CPU)|3.9|22.9|7.5|**2.9**| |Works in your browser as you read|✅|❌|❌|❌| |Remembers earlier pages and names|✅|❌|❌|❌| ¹ 33 of Yomikae's 40 skipped lines are bubbles with only "?" or "……", which it leaves as they are. *All four tools ran 214 OpenMantra pages with the same LLM (Gemma 4 12B), reasoning off, vision models on the CPU. Each ran once. Full method:* [*docs/benchmarks.md*](https://github.com/gravewhisper/yomikae/blob/main/docs/benchmarks.md)
document fetched 2026-11-17 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xxubo7/yomikae_translate_raw_manga_in_your_browser