VideoSubtitleRemoverPro v3.17.1

What's new in v3.17.1

Detection & OCR

• RapidOCR 3.x support -- lifted pip cap from <3.0.0 to <4.0.0; hardened constructor and output parsing for v3.x API changes
• PP-OCRv6 auto-selection -- PaddleOCR 3.7+ defaults to PP-OCRv6 (34.5M params, +5.1% accuracy, 5.2x CPU speedup, 50-language single model)
• Vertical text detection fix -- coordinate mapping after 90-CCW rotation was mirrored/swapped; boxes now placed correctly

Security

• VSR_SEEDVR2_CMD now uses shlex.split() to prevent argument injection via environment variable
• Crash reporter path regex extended to handle paths with spaces (prevents partial username leaks)
• Support bundle restricts collected paths to safe file extensions (.json, .md, .txt, .log)

Reliability

• Fixed preset save crash (BUILTIN_PRESETS and _load_user_presets missing imports)
• Fixed crash in queue item elapsed-time display (missing datetime import)
• Fixed CAP_PROP_FRAME_COUNT returning -1 bypassing the fallback (used max(1, ...))
• Catch TclError alongside RuntimeError when posting to Tk from worker threads
• A/B scrubber releases VideoCapture objects even if window construction fails
• JSON log handler now properly closes its file handle
• Corrupt settings.json backed up to .bak before resetting to defaults

UX

• "Set region" button now appears directly in the preview action bar
• Preview height increased from 158px to 260px
• Drop area shows accurate text when TkinterDnD2 drag-drop is unavailable
• Status badge, progress bar, and accent stripe now forward clicks in queue cards
• ModernButton.set_enabled(False) resets stale hover/pressed state
• GitHub issue templates with OS/GPU fields and feature request form

Build

• CycloneDX SBOM (sbom.cdx.json) generated in release builds
• ProPainter mode documented as TBE+LaMa hybrid (not upstream NTU S-Lab weights)
• Architecture docs refreshed for post-module-split tree

VideoSubtitleRemoverPro v3.16.0

Automated build for v3.16.0.

Highlights

• RTX 50-series (Blackwell) support. setup.py now auto-detects 50-series cards (5050–5090, RTX PRO 6000, B100/B200/GB200) and installs PyTorch from the CUDA 12.8 (cu128) index. Older cu118/cu121 pins have no sm_120 kernels and error out or fall back to CPU.
• Fixed a stale torch==2.7.0 README pin that conflicted with the torch>=2.10.0 floor (CVE-2026-24747).
• Documented TV / limited color-range handling (preserve_color_metadata, on by default).

Which download do I want?

• GPU users (especially RTX 50-series): run from source — clone the repo at this tag and follow the install steps so PyTorch lands on the right CUDA wheel (cu128 for Blackwell). The attached ZIP is a CPU-only PyInstaller build and will not use your GPU.
• CPU users / quick try: VideoSubtitleRemoverPro-v3.16.0-windows.zip is a self-contained Windows build — unzip and run VideoSubtitleRemoverPro.exe.

Note on antivirus / SmartScreen

This ZIP is an unsigned PyInstaller bundle that ships CUDA/torch and fetches model weights on first run. AV engines and SmartScreen commonly flag that pattern heuristically (large packed binary + network calls) as a false positive. Downloaded model weights are verified against vendored SHA-256 hashes. If your AV flags it, that's the heuristic — verify the ZIP hash and run from source if you prefer.

Full Changelog: v3.5.0...v3.16.0

Video Subtitle Remover Pro v3.5.0

Video Subtitle Remover Pro v3.5.0

Full rewrite from v2.0: real AI inpainting, multi-engine detection, professional DPI-responsive GUI, 20+ bug fixes.

---

AI Inpainting Engine

• Real LaMa neural inpainting via simple-lama-inpainting -- dramatically better quality than OpenCV
• STTN and ProPainter modes available with OpenCV inpainting fallback
• Image output quality preserved per-format (JPEG 95%, PNG compression 3, WebP 95%)

Multi-Engine Text Detection

• Automatic fallback chain: PaddleOCR > EasyOCR > OpenCV
• 12 languages: English, Chinese, Japanese, Korean, French, German, Spanish, Portuguese, Russian, Arabic, Hindi, Italian
• EasyOCR language code mapping for cross-engine compatibility
• Adjustable detection threshold (10-90%) in Advanced Settings
• Improved OpenCV fallback: dual-threshold bright/dark text detection with morphological merge

Professional GUI

• DPI-responsive layout -- Per-Monitor DPI V2, proportional grid columns, all widgets resize dynamically
• Unified panel design -- Input and Settings merged with subtle divider, no cluttered borders
• Dark-themed throughout -- combobox dropdowns, scrollbars, label frames, scale sliders all properly themed
• Subtitle region selector -- draw a rectangle on the first frame to define the target area
• Detection mask preview -- right-click queue item to visualize detected text regions
• Before/after comparison -- click completed items for side-by-side preview
• Video time-range -- process only a segment (start/end seconds)
• Folder input -- drag-drop folders or right-click drop zone to browse folders
• Batch progress bar with window title progress indicator
• Collapsible log panel with color-coded output and Open Log File button
• CRF quality slider in Advanced Settings
• Tooltips on key controls with full-path tooltips on filenames
• Retry Failed button to re-queue failed/cancelled items
• Double-click completed item opens the output file
• Taskbar flash + sound on batch completion
• Settings lock during processing
• Custom output directory with Browse/Reset
• Queue capped at 500 items
• Clean, professional spacing at any resolution from 900x650 to 4K

Settings & Persistence

• All settings saved to %APPDATA%\VideoSubtitleRemoverPro\settings.json
• Subtitle region, language, threshold, CRF quality, GPU selection all persist
• Crash logging to rotating log file (5MB) + MessageBox on fatal error

CLI

python -m backend.processor -i input.mp4 -o output.mp4 -m lama --lang en --crf 20 --threshold 0.4 --start 30 --end 120

Flags: --mode, --gpu, --lang, --crf, --threshold, --start, --end, --skip-detection, --fast, --no-audio

Infrastructure

• GitHub Actions CI/CD workflow for automated Windows builds
• .gitignore for model weights, venv, build artifacts
• setup.py auto-installs EasyOCR + simple-lama-inpainting + PaddleOCR
• All source files pure ASCII (zero unicode characters)
• Branch protection enabled on main

Bug Fixes (20+)

• Processing was simulated, never called real backend
• Backend path pointed to nonexistent file
• InpaintMode enum mismatch between GUI and backend
• Race conditions in queue mutations during processing
• Widget destruction during processing broke progress updates
• GPU selection never synced to config
• Duplicate files could be queued
• Output files silently overwritten (now auto-increments)
• Cancel only marked one item, remaining abandoned (now marks all)
• Elapsed timer dict iteration crash
• Settings lock disabled entire widget tree (now scoped)
• Closing window during processing left orphan threads
• nvidia-smi crash on malformed output
• EasyOCR language code mismatch (ch vs ch_sim)
• FFmpeg subprocess deadlock (added 10min timeout)
• Subtitle area not persisted in settings
• ModernButton fired on release outside bounds
• Region selector VideoCapture resource leak
• Division by zero on 0-frame video containers
• Batch progress bar stuck at 100% after completion
• Eager output directory creation (deferred to processing)

v2.0

Add files via upload