Version History
Track all updates, new features, and improvements to Thaluna Real-Time Translator.
Thaluna 3.1 Preview - Audio Translation Beta, OCR Improvements & Lens Stability
- New: Added beta Audio Translation for Real-Time mode using live loopback/system audio transcription instead of screen OCR.
- New: Added OpenRouter Speech-to-Text support with source language selection, STT model selection, and latency presets.
- Changed: Stable latency is now the recommended default for Audio Translation because it gives cleaner speech recognition than lower-latency testing modes.
- New: Added optional real-time OCR auto-crop to trim excessive empty space before OCR processing.
- Improved: Calmer OCR logging and reduced repeated identical OCR log spam.
- Improved: Lens Mode stability, Continuous Lens behavior, overlay capture exclusion, manual snapshot handling, and multi-monitor usability.
- New: Added Lens customization controls for hiding the Lens UI, Hide After timing, background opacity, and overlay readability.
- Fixed: OpenRouter requests from slow reasoning-heavy models no longer block the app or delay model/provider switching for minutes.
- New: Added a Settings warning when the selected OpenRouter model looks like a reasoning/thinking model that may be too slow for Real-Time OCR.
- Changed: Updated the default OpenRouter preset list with GPT-4.1 Nano and removed slower default presets that were less suitable for realtime translation.
- Updated: Refreshed internal Python dependencies and added audio/VAD-related components to support the new live speech pipeline and reduce known dependency vulnerabilities.
- Notes: Audio Translation is marked BETA because quality depends on source audio, selected STT model, background noise, overlapping voices, and audio device configuration.
Thaluna 3.0 Preview - Local API, Lazy Loading & Stability Pass
- New: Added Custom OpenAI-compatible Local API support for local servers such as llama.cpp, LM Studio, and similar tools.
- New: Custom Local API supports configurable base URL, model ID, and optional API key storage.
- Improved: Lazy OCR/model loading reduces startup pressure and loads heavier components only when needed.
- Improved: Better Paddle/Torch DLL initialization for packaged builds, including deferred Torch runtime loading.
- Improved: More robust model switching across built-in local models, Ollama, OpenRouter, and Custom Local API.
- Fixed: Model index preservation now ignores separator rows and restores selections by model name when dynamic model lists change.
- Fixed: OpenRouter without an API key now pauses translation instead of repeatedly throwing errors or silently switching models.
- Fixed: API keys are kept out of
config.json and stored in local per-user secrets files.
- Notes: This is a Preview build. The previous v2.9 stable build remains available on Itch as a fallback download for buyers.
🔊 Thaluna v2.9 - Real-Time TTS + Better Startup Guidance
- New: Added optional Real-Time TTS using your existing OpenRouter API key
- New: Choose whether TTS reads translated text or original OCR text
- New: Added voice selection and adjustable speech speed for Real-Time TTS
- Improved: Clear startup guidance when Thaluna is launched from non-ASCII or accented folder paths that can break bundled local model loading
- Improved: Better validation for TTS voice settings to prevent invalid custom values from being saved
- Notes: Real-Time TTS is optional and requires a configured OpenRouter API key
🔍 Thaluna v2.8 - Continuous Lens, DPI Fix & Lens Improvements
- New: Continuous Lens inside Lens Mode for automatic selected-area refresh when the image changes
- New: Dedicated Lens settings page with Continuous Lens sensitivity, minimum auto-refresh delay, anti-flicker controls, and a manual Lens OCR behavior toggle
- New: Manual Lens can optionally force a fresh OCR engine per snapshot when you want slower but potentially cleaner OCR behavior
- New: Experimental alternate capture path for Windows scaling above 100% in the standard real-time workflow
- Improved: Continuous Lens now reduces overlay flicker through an anti-flicker capture path on supported Windows setups
- Improved: Continuous Lens now handles mode switching, shortcuts, and manual snapshot conflicts more safely
- Notes: Continuous Lens and the DPI/scaling fix are opt-in features. Default behavior stays unchanged unless you enable them.
📦 Thaluna v2.7 - Stability Improvements, Backlog, GPU Fix & Transcript Saving
- Improved: Clearer startup and error messages for common Windows memory/security issues, including better handling of WinError 1455 cases
- Fixed: Internal configuration synchronization between UI, worker, and Lens Mode for more reliable provider/model consistency
- Fixed: Rare Lens Mode state mismatch after specific Settings/Cancel/model-switch sequences
- Fixed: Packaged GPU OCR runtime in Full and Demo builds now bundles required CUDA/cuDNN components correctly for the Paddle GPU backend
- Improved: Internal DLL loading order for Paddle GPU runtime and better fallback diagnostics/logging in packaged builds
- New: VN-style Translation History (Backlog) overlay for reviewing recent translated lines
- New: Automatic session transcript saving to
.txt with folder selection and save mode options
- Improved: Reduced startup overhead by loading Manga OCR only when Manga mode is used
- New: OCR capture diagnostics logging to help troubleshoot rare DPI/scaling offset issues on Windows
- Notes: This update focuses on stability, packaging reliability, and UX consistency rather than OCR/translation accuracy changes
🚀 Thaluna v2.6 - OpenRouter + Stability Improvements
- New: Added optional OpenRouter cloud translation mode
- New: Preset model IDs: Gemini 2.5 Flash Lite, GPT-OSS 120B, GPT-4o Mini, DeepSeek V3.2
- New: Custom OpenRouter model ID support
- New: Instant / Economy update behavior for cloud mode
- New: Cloud mode can reduce local GPU load in heavier games
- Fixes: Improved output-language enforcement for OpenRouter responses
- Fixes: Improved fallback behavior when cloud key is missing or invalid
- Fixes: Safer request + caching behavior to prevent cross-language mix-ups
- Security: API key stored per-user in LocalAppData, with automatic legacy migration
- Notes: OpenRouter model names are convenience presets only; Thaluna does not host or distribute third-party models
🆓 Demo: All Languages & UX Fixes
- All 30+ languages now unlocked in Demo
- Translation limit increased: 30/day (was 20)
- Auto OCR language detection when switching models
- GPU fallback to CPU if initialization fails
- Fixed VRAM leak and memory management
📦 Stable: New Languages & Critical Fixes
- 12 new languages now stable: Arabic, Hungarian, Dutch, Finnish, Danish, Norwegian,
Bulgarian, Serbian, Croatian, Slovak, Hindi, Russian
- Fixed crashes for non-ASCII usernames (Cyrillic, Arabic, Chinese)
- Experimental channel replaced with Diagnostic build
- Better error messages for missing system components
- RTX 5000: OCR on CPU recommended until framework support added
🧪 Experimental: Language Expansion
- 11 new languages: Hungarian, Dutch, Finnish, Danish, Norwegian, Bulgarian, Serbian,
Croatian, Slovak, Hindi, Russian
- Direct translation from English + two-step via English for JA/ZH/KO
- File logging system (logs at %LOCALAPPDATA%\Thaluna\logs\)
- CMD console window removed from startup
- Auto log rotation (max 3 files)
Experimental Channel & Arabic Support
- Experimental build channel for early features
- Arabic language support
- Arabic-compatible font with automatic switching
- Improved RTL text handling
Translation Stability Hotfix
- Improved OCR and translation synchronization
- Correct OCR language selection on startup
- Fixed UI layout issues when resizing
- General stability and consistency fixes
New Languages & UI Refresh
- New languages: Greek, Indonesian, Swedish, Thai, Vietnamese
- Japanese, Korean, and Chinese available as targets
- Redesigned language selection menu
- UI polish and configuration stability improvements
The Customization Update
- Presets system for saving and switching layouts
- Customizable UI colors
- Global shortcuts system
- Refreshed settings menu
- Automatic GPU to CPU fallback
OCR Startup Models Fix
- Faster first launch with pre-bundled OCR components
- No changes to v2.1 features
The Manga & Stability Update
- Snapshot Mode: move, resize, and remove translation boxes
- Vertical Japanese support for manga speech bubbles
- New OCR language selector for accurate detection
- OCR on GPU option with automatic CPU fallback
- Improved model switching and memory cleanup
Major Update: Lens Mode & Manga Mode
- Lens Mode for full-screen or region translation
- Manga Mode optimized for speech bubbles
- New language: Turkish
- Customization for subtitle and box colors
- Stability fixes for CPU and GPU switching
Now Supports More Languages
- New built-in targets: Portuguese, Italian, Romanian, Czech, Ukrainian
- Lightweight models for low-end PCs
- Ollama support for broader language coverage
