starred/WatermarkRemover-AI
1
0 Fork 0
mirror of https://github.com/D-Ogi/WatermarkRemover-AI.git synced 2026-10-21 06:10:19 -05:00
Code Issues Packages Projects Releases 1 Wiki Activity

1 Release 1 Tag

RSS Feed
  • v0.68.0 553b9bcfc2
    Compare

    0.68.0 — Windows desktop (CPU + NVIDIA GPU) Stable

    dean released this 2026-09-16 07:59:44 -05:00 | 10 commits to main since this release

    📅 Originally published on GitHub: Wed, 16 Sep 2026 13:22:25 GMT
    🏷️ Git tag created: Wed, 16 Sep 2026 12:59:44 GMT

    Windows desktop downloads

    Download a complete portable archive below, extract it into a writable folder, then open WatermarkRemover-AI.exe. Keep the EXE alongside its python, ui and other bundled folders. No system Python installation or administrator access is required.

    Recommended: CPU + NVIDIA GPU in one package

    Download Windows CPU + NVIDIA GPU (2 in 1) — about 1.68 GiB. Supports both CPU and compatible NVIDIA GPUs. The application automatically chooses CUDA when available and CPU otherwise. Extract with 7-Zip.

    Optional smaller CPU-only download — about 470 MiB. Leaves out the NVIDIA runtime libraries and always processes on CPU.

    The main package uses 7z compression to fit GitHub's per-file asset limit. Both packages contain the native EXE launcher and their own Python/Qt runtime. These are portable packages; no installer is needed.

    First launch

    Open Models, then Download / Retry to prepare Florence-2 and LaMA (about 1.8 GB). Downloads show progress and can be retried; verified models are reused locally. Settings, logs and model caches are stored in the extracted package's data directory. Internet access is needed for initial model downloads, not for the bundled interface assets.

    FFmpeg is an optional external dependency for preserving video audio. Without it, output videos are silent. The executable is not code-signed.

    Changes

    • Standalone LaMA runtime removes the IOPaint dependency.
    • Local desktop assets and a native Windows launcher support the packaged application.
    • Model preparation reports download progress, readiness and retryable failures.
    • Theme fixes prevent sidebar horizontal overflow and improve model-dialog text contrast.
    • Updated Japanese and Chinese satire; new opt-in CorpSlop and English (Israeli PR) interface variants.

    Verification

    Each downloadable archive was extracted into a fresh directory and checked file-for-file against its build folder. The extracted EXE passed the real Qt desktop/JavaScript bridge check across 64 theme/language combinations at 800×600; minimum measured model-panel text contrast was 6.7:1. CI passed on Windows, Linux and macOS for the included changes.

    Both the final CPU package and the 2-in-1 package with CUDA devices hidden passed real-model CPU image processing, detection preview, cancellation and missing-input failure checks. The 2-in-1 fallback test used Torch 2.14.0+cu126 with CUDA availability false. The final 2-in-1 package also passed verified LaMA inference on an NVIDIA RTX 4090 using a 96×64 synthetic input, preserving unmasked pixels. Full Florence-2 + LaMA GPU inference was not revalidated for this release under the workstation's current GPU load. Hardware compatibility beyond the tested environment is not implied.

    Use SHA256SUMS.txt to verify archive integrity. Each archive also includes build-info.json, installed-packages.txt and third-party license notices.

    Source commit: 553b9bcfc29ebfc9e7b2f5779091379be2db8ba1.

    Release metadata was refreshed after an author-email privacy cleanup. Only the source commit in build-info.json changed; application and bundled runtime files are unchanged. Archive checksums have been updated accordingly.

    Downloads