• untaint.content.ts: New module for safer canvas manipulation, preserving DOM integrity when reading pixel data from cross-origin or tainted canvases.
• Re-Translate Text from Edit Mode: Beside Editing the Text in Typeset Mode, you can now retranslate it via 1-click button for each detected text box.
• Sunday Webry Adapter: Updated site adapter with corrected selectors and proper event handling for the canvas-based reader used by the platform.
• CJK Text Wrapping: Improved line-wrap algorithm for CJK characters in canvas-rendered content — prevents mid-character breaks and overflow in translated bubbles.
• Canvas-to-JPEG Export: Enhanced export utilities with improved canvas-to-JPEG conversion for better output quality and compatibility.

• Model Select Dropdown: Firefox-specific adjustments to model selection dropdown behavior to prevent display and interaction regressions unique to Firefox's rendering pipeline.
• Offscreen Document Messaging: Optimized background communication patterns for cross-origin contexts, reducing message-passing overhead in Firefox's stricter extension model.

• Translation Registry Concurrency: Refined concurrency handling in content scripts for more stable multi-page translation operations without race conditions.
• Overlay Hover Mapping: Enhanced positioning and mapping logic for floating trigger components — hover zones now align correctly across zoomed and scaled page layouts.
• Event Propagation Cleanup: Proper teardown of overlay events prevents double-triggering in content scripts during page navigation or re-injection.
• Bubble Editor Retranslation: Improved retranslation flow from the bubble editor with toast feedback confirming the operation completed.

• Streamlined OCR Interfaces: Unified error handling across Paddle, Pororo, and Manga-OCR engines for consistent failure reporting and fallback behavior.
• LLM Model Persistence: LLM model selections are now correctly saved during the setup wizard flow and restored on next open.
• GPU Denylist Tests: Strengthened test coverage for GPU compatibility verification documentation.
• Detection Concurrency Controls: Upgraded detection main module with refined concurrency guards to prevent overlapping inference calls.

• Debug Panel: Improved debug panel visibility and added new debugging hooks for canvas pipeline inspection.

• Lighter Qwen3.5-4B Build: The Firefox default GGUF switched from IQ4_XS (~2.3 GB) to Q3_K_M (~2.1 GB), cutting VRAM and memory footprint. Existing installs with the old model ID will need to re-download or cleared.
• WebGPU Requirement Guard: Firefox local translation and site-rule generation now detect the WASM-CPU fallback upfront and fail with actionable guidance — enable dom.webgpu.enabled and javascript.options.wasm_js_promise_integration in about:config, or switch to API / Gemini mode — instead of silently crawling on CPU.
• Crash-Proof Wllama Runtime*: Crashed or aborted llama.cpp runtimes are now torn down and rebuilt automatically — fresh-instance CPU retries on load, instance disposal on weight-load timeouts, and singleton resets on mid-generation crashes — so one crash can no longer poison every later translation.
• 32-bit Memory-Safe Context: KV cache context reduced from 4096, keeping 4B GGUF loads within Firefox's 32-bit WASM memory ceiling.
• Generous Local Timeout: Local image translate requests now get 8 minutes (up from 3), preventing premature failures during slow first-run weight loading.

• Two-Way Cache Probing: The cached-model list now merges newly verified models and prunes vanished ones, so Model Storage and dropdowns always reflect reality on disk.
• Selectable Uncached Models: The model dropdown no longer greys out downloads you don't have yet — pick one and its download starts immediately.

• Smart Flipping: Dropdowns near the bottom of the viewport open upward and clamp their height to the available space.
• Scroll-Following: Scrolling the popup/sidebar repositions the open list instead of snapping it shut (it only closes once the trigger leaves the screen).

• Ollama CORS Guide: Added a Windows-specific reminder: background Ollama must be fully quit from the system tray and relaunched to pick up OLLAMA_ORIGINS changes.

• 2GB Memory Limit Fix (Issue #6): Switched GGUF model streaming from Uint8Array accumulation to Blob streaming, eliminating Firefox's 2GB ArrayBuffer allocation crash during large model downloads.
• Accurate Cache Sizing: Cache calculations now fall back to Blob.size when content-length headers are absent, ensuring accurate download tracking and storage reporting.
• Chunk Caching: Paused downloads reliably resume from the last completed chunk without re-downloading from scratch.

• 8 Pre-Configured Local Runners: Instant setup for Ollama, LM Studio, JAN, Unsloth Desktop, llama.cpp, KoboldCpp, LocalAI, and Cherry Studio.
• Auto-Config & Guides: Selecting a runner auto-fills host, port, schema, and paths, with collapsible step-by-step setup guides (CORS, tokens, endpoints).
• Custom Mode Retained: Custom configurations and endpoints remain fully supported and editable.

• Sliding-Window Tiling: RT-DETR now tiles tall pages (>1500px) into overlapping windows to handle 10,000+ pixel webtoons/manhwa without memory exhaustion.
• Seam Deduplication: IoU containment fusion merges overlapping bubble boxes at tile boundaries to eliminate duplicate translations.
• Aspect Ratio Accuracy: Fixed detection preprocessing to prevent image squashing on non-square crops.

• Secure Key Transport: Gemini API key moved from URL query parameters to the x-goog-api-key HTTP header, preventing exposure in proxy and server logs.
• Local Mode Renaming: Renamed "WebGPU Mode" to "Local Mode" across UI to honestly reflect CPU WASM fallback stages.
• Inpaint Optimization: Inpaint time measured per cluster; finished neural LaMa patches preserved during ladder fallback without re-cleaning.
• OCR Batch Alignment: Standardized slotIndex positional tracking across PaddleOCR and Pororo engines.

• Pause & Resume: Large GGUF language model downloads (Tiny Aya, Qwen) in the Setup Wizard can now be paused and resumed on demand using HTTP Range requests.
• Dedicated Stop Button: Added a Stop action that cleanly cancels active downloads and discards retained in-memory chunk caches to allow restarting fresh from 0%.
• Memory Ingestion Guard: Pause and Stop controls are automatically disabled once the stream reaches 100% while weights are assembling and loading into memory.

• LaMa Inpainting Guard: If Quality inpainting is chosen without weights present, LMT smoothly falls back to the Fast Telea ladder without throwing errors.
• FIFO Inference Runner: Inference tasks now use isolated dead-letter queuing to ensure a single failed task never locks the pipeline.
• OCR & Overlay Enhancements: Fixed Korean Pororo line slicing, added dedicated box move handles, and resolved session disposal leaks.
• Opt-in Onboarding: All model weights across the setup wizard and settings are strictly opt-in with live progress and status indicators.

Libre Manga Translator v1.0.1 brings dedicated Korean OCR support (Pororo OCR) for manhwa/webtoons, an interactive Multi-Layer Visual Canvas Editor with inpainting patch/restore tools, an Always Inpaint override, scan caching improvements, and significant translation stability enhancements.

Everything continues to run 100% on-device in WebGPU/WASM mode by default, with Gemini cloud and self-hosted API Mode as flexible alternatives. Zero telemetry, zero external tracking, zero intermediaries.

• Dedicated Korean OCR (Pororo): Say goodbye to missed text in Korean manhwa and webtoons. We've integrated Kakao Brain's Pororo OCR (TPS-VGG-BiLSTM ONNX), optimized specifically for vertical and horizontal Korean typography.
• Multi-Layer Visual Canvas Editor: You can now toggle seamlessly between Raw Scan, Cleaned Scan (inpainted background), and Result Scan (translated typesetting) directly on the reader page.
• Inpaint Patch & Restore Tools: Need to clean a missed speech bubble or restore original background art? Use the built-in canvas mark tools to paint additional inpaint patches or restore raw artwork without re-translating.
• Always Inpaint Mode: Force inpainting across all detected speech bubbles even when conservative quality heuristics decline, ensuring no leftover untranslated text remains on complex backgrounds.
• Contextual In-Place Styling: Select any translated bubble to immediately tweak font family, font size, line height, text alignment, and color via an in-canvas floating toolbar.

A full rework of the beta series, rebuilt around one goal: reading manga should feel seamless.

The full pipeline - detection, script gate, OCR, translation, inpainting, rendering - runs on-device in WebGPU Mode, with Gemini cloud and self-hosted API Mode as opt-in alternatives. Zero telemetry, zero accounts, zero proxy in-between.

🖱 Interaction: Pill, Hub & Auto-Translate
1. Floating Translate pill. Hover any page image (≥ 260px) and a pill appears at its top-left - right where your eyes already are. No context menus, no toolbar hunting, no leaving the page. After translation it stays put as an action hub: fix a line, nudge a box, compare with the original, export the page.
2. Skip bbox refining jumps straight from detection to translation for clean pages; Auto-translate turns whole chapters hands-free, rendering each page as it scrolls into view, one at a time so nothing stutters.
3. A strict status registry dedupes repeat triggers across hover, auto-translate, and popup actions - the same page never translates twice.

🔒 Privacy & Security Core
1. Zero telemetry, permanently. All telemetry pathways, data-sharing toggles, and reporting endpoints are deleted - not disabled. WebGPU Mode transmits zero bytes anywhere.
2. SSRF-gated image proxy. The background image proxy allowlists http(s) only and rejects credentials, loopback, private/link-local IPs, and internal schemes.
3. Strict translation schemas. Every backend (WebLLM, server, Gemini) validates translations[] type and exact bbox-count match before anything is painted.
4. SHA-256 model verification. Downloads are checksummed against a known-hash manifest; mismatches evict the cache entry and abort.

🔍 Detection & OCR
1. RT-DETR default. Comic Bubble Detector (bubbles + free text) is now the default; Comic Text Detector with pixel-mask seeding stays selectable. YOLO26 models were removed - old stored selections auto-migrate to the new default.
2. 11 PaddleOCR language packs. Latin + Chinese/Japanese ship during onboarding; the rest download on first encounter while the overlay names the model. The script gate auto-switches packs per page under Auto-Detect.
3. Three OCR engines. PaddleOCR multilingual (default), PP-OCRv6 Manga (lightweight Japanese fine-tune), Manga-OCR (Japanese flagship for vertical text).
4. Script gate. A 3.7 MB on-device model holds back sound effects and wrong-language lettering; every hold-back keeps its recognized text and offers one-click Translate anyway.

🎨 Inpainting: Fast & Quality
1. Fast ladder (default, model-free). Fitted text-shaped masks through planar fill → bilateral denoise → Telea, lightest passing rung wins, decline-gated with rollback.
2. Quality mode (opt-in). LaMa neural redraw first per region (~207 MB one-time download), automatic Fast fallback wherever LaMa declines - a strict superset of Fast.
3. Declined regions are left pixel-identical and flagged, never smeared.