v0.2.6 - Snap Detection Fix, Standalone Build, Veo Demo

Snap Detection Improvements:

• Reduced guided_detect coarse scale step from 8 to 4 for finer resolution
• Replaced sqrt with cbrt size penalty — less over-penalization of small
  watermarks (~28px preview tier from Gemini web UI)
• Lowered CLI snap min_size from 32 to 16 (both fallback-region and
  explicit-region snap call sites)
• Added snap_min_size state field and Min Size slider (8-64px) to GUI
  Custom mode alongside existing Max Size slider
• Standard 48x48 (65%→72%) and 96x96 (91%) snap regression-free
• Preview 28x28 watermarks now correctly detected at 51% (was 32x32 at 41%)

Standalone Build Fix:

• Disabled OpenMP to eliminate VCOMP140.DLL runtime dependency
• MSVC OpenMP runtime (vcomp140.dll) cannot be statically linked
• Vulkan GPU inference path completely unaffected
• CPU fallback path runs single-threaded (still functional, slightly slower)
• Result: true zero-dependency standalone executable on all platforms

Veo Video Watermark Remover (Demo):

• New companion project for removing Google Veo video watermarks
• Uses the same reverse alpha blending engine as GeminiWatermarkTool
• Demo build available: https://github.com/allenk/VeoWatermarkRemover

Documentation:

• Pinned Veo demo announcement in README
• Added gwt-integrations (Claude Code Skill + MCP Server) to README
• Updated project badges and author notice

v0.2.5 - AI Denoise + Advanced CLI

AI Denoise (FDnCNN + NCNN Vulkan GPU):

• FDnCNN model (1.3 MB FP16) embedded in executable via ncnn2mem
• NCNN static build with Vulkan GPU acceleration, CPU fallback
• Gradient-masked blending: only repairs edge artifacts, preserves background
• AI Denoise set as default inpaint method (strength=120%, sigma=50)
• New UI: sigma slider, GPU device info, strength up to 300%
• Adjustable Snap search range (32-320px) for resized watermarks

Advanced CLI:

• --region for explicit watermark position (absolute or corner-relative)
• --fallback-region for search area when standard detection fails
• --snap with --snap-threshold (default 60%) to reject false positives
• --denoise ai/ns/telea/soft/off with --sigma, --strength, --radius
• Smart routing: --fallback-region + --snap skips standard detection
• Fully backward-compatible: no new flags = identical to previous behavior

Cross-platform CI with NCNN submodule on Windows/Linux/macOS.

New dependencies: NCNN (submodule, BSD-3-Clause), volk (vcpkg, MIT)
AI model: FDnCNN Color from KAIR (MIT License)
Build flag: ENABLE_AI_DENOISE=ON (default in all presets)

v0.2.4 — File Dialog Crash Fix & D3D11 Resize Jitter

Fixes:

• Fix crash when opening Save/Open file dialog (ImGui NewFrame assert)
  · Win32 modal dialog loop triggers SDL EXPOSED events
  · Event watch callback re-entered mid-frame → double NewFrame()
  · Added reentrant rendering guard to both callback and main loop

• Fix D3D11 image preview jitter during horizontal window resize
  · Remove Flush() CPU stall before ResizeBuffers
  · Use event-provided pixel dimensions (not stale SDL query)
  · Skip vsync on resize frames to push content before DWM composites
  · Only process PIXEL_SIZE_CHANGED (correct on HiDPI)

v0.2.3 — Software Inpainting, Guided Detection & Rendering Fixes

New Features:

• Multi-scale guided watermark detection (Snap Engine)
  · Coarse-to-fine NCC template matching in user-drawn search region
  · Variable watermark sizes (24-160px) with scale step refinement
  · Auto-snap on draw/move release, manual re-snap via [Snap] button
  · Dual-box overlay: blue search region + green/yellow/red snap result
  · Process uses snap result (effective_rect) when available

• Software Inpainting cleanup for residual artifacts
  · Three methods: NS (Navier-Stokes), TELEA, Soft Inpaint
  · Gradient-weighted masks derived from watermark alpha channel
  · Configurable strength, radius, and method in Custom mode UI
  · Addresses images degraded by post-watermark resize/recompression

Rendering:

• Fix window resize flickering (DXGI_SCALING_NONE replaces STRETCH)
• Fix D3D11 image preview jitter during horizontal resize
  · Remove Flush() stall, skip vsync on resize frames
  · Use pixel-accurate event dimensions for ResizeBuffers
• Only process SDL_EVENT_WINDOW_PIXEL_SIZE_CHANGED (HiDPI correct)

v0.2.2 - macOS Compatibility Release

Fixes macOS-specific issues with window sizing, Retina rendering, and
first-run security, plus preview enhancements across all platforms.

• Fix fullscreen clipping on small displays (e.g. MacBook 1280x800)

  • Root cause: kMinHeight (888) exceeded macOS usable screen height (708)
  • Minimum window size now adapts to 95% of display usable area

• Retina font clarity: rasterize font atlas at 2x framebuffer resolution
  with FontGlobalScale compensation — pixel-perfect text on HiDPI

• Ad-hoc code signing in CI — eliminates "app is damaged" error

  • Users can now right-click → Open instead of using xattr command

• Enable GUI build for both x64 and arm64 architectures

• Zoom-to-cursor: mouse wheel zooms toward cursor position

• Zoom buttons and shortcuts zoom toward viewport center

• Outlined text labels (4-directional dark outline) for readability
  on both light and dark image backgrounds

• Hold C to temporarily hide all region overlays for clean inspection

• Updated shortcut table in sidebar and README

• Fix toolbar separator artifact (stray horizontal line between buttons)
  on all platforms — move Separator to toolbar top as menu/toolbar divider

• Rewrite DXGI initialization to cooperative mode

  • Factory-first flow, no exclusive flags, no mode switching
  • Resolves mutual exclusion with MPC-HC/BE, games, and other D3D11 apps

• Lightweight is_available() with zero device creation side effects

• SetMaximumFrameLatency(1) to reduce GPU present queue footprint

• Add --banner/--no-banner flags with TTY auto-detection

• Auto-hide ASCII banner when piped or redirected (AI agent friendly)

• Add "First Run — OS Security Prompts" guide (macOS / Windows / Linux)

• Update keyboard shortcuts in README

v0.2.1 - Windows Compatibility Release

This release significantly improves Windows compatibility, especially for
virtualized environments and CJK (Chinese/Japanese/Korean) users.

• Add D3D11 as the preferred render backend on Windows
• Automatic WARP (software rasterizer) fallback when no GPU available
• Works in Hyper-V, Docker, Remote Desktop, and other virtualized environments
• Cooperative DXGI mode: no exclusive flags, no mode switching
  • Factory-first init: CreateDXGIFactory1 → EnumAdapters → CreateDevice
  • Lightweight is_available() with zero device creation side effects
  • Resolves mutual exclusion with MPC-HC/BE, Forza Horizon 5, and other D3D11 apps
  • SetMaximumFrameLatency(1) to reduce GPU present queue footprint
  • ClearState() + Flush() for clean shutdown
• Use --backend=opengl to force OpenGL if needed

Tested:

• NVIDIA RTX PRO 6000 (Hardware D3D11) alongside MPC-BE and Forza Horizon 5

• Microsoft Basic Render Driver (WARP fallback in Hyper-V)

• Zoom-to-cursor: mouse wheel zooms toward cursor position

• Zoom buttons/shortcuts zoom toward viewport center

• Outlined text labels (4-dir 1px dark outline) for readability on any background

• Hold C to temporarily hide all region overlays for clean result inspection

• Updated shortcut table: C (hold), Ctrl +/-, Ctrl 0, Scroll

• Add --banner/--no-banner flags with TTY auto-detection

• Auto-hide ASCII banner when piped or redirected (AI agent friendly)

• --no-banner for scripted/automated usage, --banner to override

• Fix crash when drag-dropping files with CJK characters (中文/日本語/한국어)

• Add path_from_utf8() for correct SDL UTF-8 to filesystem path conversion

• Improve CLI error messages with guidance for CJK path issues

• Centralize console UTF-8 setup for both CLI and GUI modes

• Clamp initial window size to 95% of usable screen area

• Prevent window from being larger than display on small screens

• Disable imgui.ini generation (unused with viewport-based layout)

• Add WIL (Windows Implementation Libraries) for D3D11 COM management

• Add imgui[dx11-binding] for D3D11 ImGui integration

v0.2.0 - GUI Application & Watermark Detection

• Full graphical desktop application with ImGui + SDL3

  • Drag & drop single image or multiple files for batch processing
  • Drag & drop entire folder to load all supported images
  • Real-time before/after comparison, zoom, pan, fit-to-window
  • Custom watermark mode: draw region interactively with 8-point anchors, WASD fine-tuning
  • Keyboard shortcuts: X (process), V (compare), Z (revert)
  • CJK font support (Chinese/Japanese/Korean filename display)
  • Native file dialogs with Linux fallback (zenity/kdialog)

• Three-stage NCC detection algorithm — inspired by @dannycreations (#13)

  • Stage 1: Spatial NCC with alpha map correlation (50% weight)
  • Stage 2: Gradient NCC via Sobel edge matching (30% weight)
  • Stage 3: Statistical variance analysis (20% weight)
  • Circuit breaker at 0.25 to short-circuit obvious non-matches

• Confidence scoring with configurable threshold (default: 25%)

• CLI: --force to skip detection, --threshold for custom value

• GUI: interactive threshold slider (0-100%, 5% steps)

• Thumbnail atlas preview (4-column grid, 220px cells, filename labels)

• Status overlays per image: OK / SKIP / FAIL

• Confirmation dialog before overwriting originals

• Progress bar with scrollable result log

• Thumbnails refresh after completion

• Reorganized source into src/core/, src/cli/, src/gui/

• Core engine shared between CLI and GUI

• Theme constants extracted to �atch_theme namespace in style.hpp

• C++ standard updated from C++17 to C++20

• @dannycreations - Watermark detection concept (#13)

Commits since v0.1.3:

• 278492b feat(gui): batch processing UX improvements and directory drop support
• 2da7a83 feat(detection,gui): three-stage NCC detection, CLI batch processing, and UI enhancements
• b81599b feat(gui): complete UI implementation with CJK support, live resize and UX improvements
• d21b69d feat(gui): implement ImGui scrollbar-based image panning and fix stability issues
• 12c865c refactor!: reorganize project structure and relocate files

v0.1.3 - Multiple files support & UTF-8 path fix

• Multiple files support in simple mode (#12) - Thanks @dannycreations!

  • Drag & drop multiple files onto the executable
  • Process multiple files with single command: GeminiWatermarkTool file1.jpg file2.png file3.webp
  • Summary reporting shows success/failure counts after batch processing

• Windows UTF-8 path display fix

  • Fixed garbled characters when processing files with non-ASCII names (e.g., Chinese, Japanese, Korean)
  • Paths like "複製.png" now display correctly instead of " ƻs.png"

• Refactored CLI with unified ProcessResult tracking

• Extracted process_single helper to reduce code duplication

• Consolidated result printing across simple and batch modes

• @dannycreations - Multiple files support (#12)

Commits:

• 5d3d473 feat(cli): support multiple files in simple mode (#12)
• c9797eb feat(core): add path_formatter for Windows UTF-8 path display

v0.1.2 - Cross-Platform Release

This release brings full cross-platform support with pre-built binaries for Windows, Linux, and macOS.

• Cross-platform support: Windows, Linux, macOS

• macOS Universal Binary (Intel + Apple Silicon)

• CMakePresets.json for standardized builds

• GitHub Actions CI/CD for automated releases

• Windows: Windows 10/11 x64

• Linux: x64, glibc 2.35+ (Ubuntu 22.04+, Debian 12+)

• macOS: macOS 11.0+ (Intel or Apple Silicon)

For older Linux systems (Ubuntu 20.04+), build from source using:
cmake --preset linux-x64-Release
cmake --build --preset linux-x64-Release

See README.md for full build instructions.

🎉 Open Source Release

This release marks the full open source release of Gemini Watermark Tool. The complete source code is now available!

✨ Features

• Reverse Alpha Blending - Mathematically accurate watermark removal
• Auto Size Detection - Automatically detects 48×48 or 96×96 watermark
• Simple Mode - Just GeminiWatermarkTool.exe image.jpg for in-place edit
• Batch Processing - Process entire directories at once
• Standalone Binary - Single .exe, no dependencies

🔧 Build Modes

📐 Watermark Size Rules

🐛 Bug Fixes

• Fixed: 1024×1024 images now correctly use Small (48×48) watermark
• Fixed: --force-small / --force-large flags now work correctly
• Fixed: Output quality improved (JPEG 100, WebP lossless)

📦 Downloads

🛠️ Build Requirements

• CMake 3.20+
• C++20 compiler (MSVC 2022, GCC 11+, Clang 14+)
• vcpkg (for dependencies)

📋 Dependencies

• OpenCV 4.x
• CLI11
• spdlog
• fmt

⚠️ Disclaimer

Always backup original images before processing. See README for full disclaimer.

Full Changelog: https://github.com/allenk/GeminiWatermarkTool/commits/v0.1.1