Skip to content

1.1.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 17 Aug 20:53
· 7 commits to main since this release
v1.1.2
00f4aee

Assets removed — superseded by v1.1.3. This release's binaries worked correctly and have been deleted only because v1.1.3 supersedes it entirely. Get v1.1.3 instead.

Latent Tools — v1.1.2

Local-first Windows desktop app for bulk image-dataset prep: AI watermark
removal, format conversion, and uncensored image captioning — all running
entirely on-device on your own GPU.

What's new in v1.1.2

  • Fix: packaged sidecar crashed on startup, every release through v1.1.1.
    Nobody had actually run the packaged .exe and checked GPU status since
    v1.0.0 first shipped — CI only ran the unpackaged test suite, which never
    exercises the frozen binary. Two independent bugs, fixed here:
    • uvicorn.run("app.main:app", ...) used the "module:attr" string form,
      invisible to PyInstaller's static analysis, so it silently dropped the
      app package from the bundle and the packaged sidecar crashed instantly
      with ModuleNotFoundError: No module named 'app'.
    • Past that, the sidecar crashed with OSError: [WinError 1114] ... c10.dll.
      The real fault (only visible via Windows Event Viewer) was an access
      violation inside a stale, PyInstaller-auto-bundled msvcp140.dll that
      shadowed the correct, much newer system copy — torch's c10.dll (built
      against a newer CRT) crashed calling into it.
    • Full technical detail is in HANDOVER.md in the repo.
  • New app icon — LT monogram replaces the previous mark across the
    installer, titlebar, and system tray.

Highlights (full feature set)

  • Watermark removal — Florence-2 open-vocabulary detection finds
    watermarks/logos/text, LaMa (IOPaint) inpaints them out. Manual mask
    brush/eraser editing with undo/redo if the automatic detection needs a
    nudge.
  • Format conversion — JPEG / PNG / WEBP export with quality, lossless,
    compression-level, background-flatten color, and metadata-retention
    controls. Export presets for LoRA / Archive / Web, plus custom
    localStorage-backed presets.
  • Uncensored image captioning — Qwen2-VL-2B / Qwen2-VL-7B-Instruct, or
    point it at your own local model folder. Custom system prompts and
    trigger-word support.
  • Bulk dataset processing — folder-in, folder-out batch pipeline built
    for preparing training datasets (e.g. LoRA), not just one-off cleanup.
    Single round-trip /process pipeline (normalize → detect → inpaint →
    caption → convert) for throughput.
  • Single Image Editor — Detect → Remove → Caption stepper with a
    zoomable/pannable canvas mask overlay.
  • Live GPU telemetry — real-time GPU name, VRAM usage, and temperature
    in the titlebar and sidebar.
  • Runs fully local — no cloud calls; a Python (FastAPI) sidecar talks
    to the Electron app over 127.0.0.1 only.

Requirements

  • Windows 10/11 (64-bit) — no macOS/Linux build today.
  • A dedicated, CUDA-capable NVIDIA GPU. There is no CPU fallback.
    8GB+ VRAM recommended; 16GB+ if you plan to run the 7B captioning model.

Installation

Download and run either:

  • Latent-Tools-Setup-*.exe — NSIS installer
  • Latent-Tools-*.exe (no "Setup" in the name) — no-install portable build

Known limitations

  • Windows-only.
  • First run downloads model weights (Florence-2, LaMa, Qwen2-VL) from
    Hugging Face — expect a delay and disk usage on first launch.
  • No auto-update mechanism yet — check the Releases page for new versions.
  • Each packaged .exe was ~616MB — that's the correct cost of bundling
    PyTorch, CUDA, and the model stack (v1.1.1's much smaller build looked
    efficient but was actually broken — see v1.1.3's notes). Not a bug.

Full Changelog: v1.1.1...v1.1.2

Assets 2
Loading