LIBRE MANGA
TRANSLATOR

Raw scan in.
Translated page out.

LMT erases original text with a fitted text-shaped mask and repaints your language inside the same bubble — screentone and artwork survive untouched. Keep scrolling to wipe between the raw page and the result.

• Auto inpaint ladder — planar fill → bilateral denoise → LaMa redraw (opt-in) → pure-JS Telea, decline-gated
• Every rung scored by one quality check — a region nothing can clean is left pixel-identical and flagged declined
• Minimal ghost rectangles, no flat patches, no halos

Three modes.
One pipeline.

Libre — you choose where your data goes. Default to fully on-device, take the cloud when your GPU is weak, or keep total control with your own server. Toggle back to the original image any time.

• Dual schema support: OpenAI json_object and LM Studio experimental payloads
• Test Connection probe verifies host, schema and model before you translate
• Model Storage lists every weight with size — per-model delete and full cache clear

Heavy inference,
off the main thread.

A YOLO26-Nano bubble detector runs in a dedicated offscreen document — the page you're reading never drops a frame. Boxes are merged, speckle-dropped, and numbered in right-to-left reading order. Small, RT-DETR and ComicTextDetector (GPL-3.0, on-demand) are selectable in Settings › Detection.

• Language gate — a 3.7 MB on-device script-ID model holds back sound effects and artwork lettering
• Series context (title, summary, glossary, last 5 pages) keeps names consistent — results keyed by series + chapter + page-index + image hash

The box editor
is honest.

Detected boxes are numbered in manga reading order. Drag a missed bubble into place, resize the ones the detector got wrong, undo and redo freely — before a single token is spent on translation.

• Right-to-left, top-to-bottom numbering with panel-aware band grouping
• Every correction can train the detector — opt-in, anonymous, off by default
• Overlay isolation + natural-vs-rendered scaling fix — editor keys never fire the host reader

Read the OCR.
Fix the line.

Open the Edit modal to see the extracted text and the translation side by side, per bubble. Tweak a line, hit Apply — it re-renders on the canvas from the cached inpainted base, without re-running inpainting.

• Gate-held regions keep their recognized text and show a one-click Translate anyway
• Toggle original vs translated at any moment to compare

Your site,
your rule.

LMT matches the current URL against community adapters to pull series name, chapter ID and page index. If your site isn't covered, the AI rule generator drafts one from your active model — or write it by hand in about three minutes.

• Drop a file in src/lib/adapters/ — auto-bundled at build, no registry edits
• Floating on-page sidebar: switch mode mid-chapter without leaving the page
• Anti-hotlink bypass + magic-byte decode — JPEG, PNG, WebP, GIF, AVIF

Three steps
to your first page.

No cloud subscription, no account, no credit card. In WebGPU mode the whole pipeline works offline after the initial model download. Same-device API Mode (Ollama / LM Studio on localhost) is offline-capable too.

• Right-to-left box sorting is already done for you
• Results cache per page — re-opening a translated chapter is instant
• Self-host on the same device — Ollama 127.0.0.1:11434 or LM Studio 127.0.0.1:1234, no internet needed after setup

Built in the open.
Shipped honestly.

LMT releases are intentionally infrequent while the pipeline stabilizes. This is what's actually been delivered, what's being worked on right now, and what's next — not a wishlist.

• OCR can fail on Chinese/Korean manhua & manhwa — actively investigated
• Gate false negatives - heavily stylized or mixed-script lettering can still be read as the wrong script. The language gate marks every hold-back and offers Translate anyway, so it never silently drops a bubble.

Built on open work.
Shipped MIT.

LMT is MIT — built upon ComicTL, with API Mode roots and ladder/gate concepts studied, then rewritten independently.

• ComicTL by Ketut Shridhara — YOLO pipeline, PaddleOCR mapping, WebLLM Qwen3, editor, context, adapters, Gemini path
• Local Manga Translator PoC — LMT acronym origin, self-hosted servers, dual schema, Test Connection, debug logs
• Manga Cleaner Ladder inpainting concepts (GPL-3.0, KoMiQ) — fitted masks, ladder + gate, denoise — independent TS rewrite, no copied code
• Weights: YOLO26 MIT · RT-DETR Apache-2.0 · ComicTextDetector GPL-3.0 (on-demand) · OSD Apache-2.0 · Paddle/Manga-OCR Apache-2.0 · LaMa MIT

100% on-device by default · privacy.md ↗