v1.1.0

Desktop app 1.1.0

Brings the new automatic removal to the desktop app: isnet with the ormbg
second opinion, run natively, and the same backdrop refinement as the web.
The interface's copy of the model is left out of the installer, since the
app serves models from its own bundled folder.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013VzLN37VFjdNm7oTdMq5ZM

v1.0.1

Desktop app 1.0.1

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

v1.0.0

Add a desktop app for macOS and Windows

The desktop build runs automatic removal through the native ONNX Runtime
instead of WebAssembly, about eight times faster, with the model bundled so
it works offline. It reuses the web interface unchanged.

- desktop/: Electron shell, inference in a utility process, self-check
- lib/engine/native.ts: the seam the editor uses when the bridge is present
- header button that detects the visitor's system and offers both installers
- release workflow that builds and attaches both installers to a tag

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Models 1

Model files used by the app and the desktop builds. Not an app release.

isnet-general-use-q8.onnx: the automatic background remover.

• Weights: isnet-general-use from the DIS authors (Qin et al., Highly Accurate Dichotomous Image Segmentation, ECCV 2022), https://github.com/xuebinqin/DIS, Apache-2.0. The DIS5K training dataset has its own terms of use.
• ONNX conversion: rembg, https://github.com/danielgatis/rembg, MIT.
• 8-bit weight quantization: onnxruntime.quantization.quantize_dynamic (QUInt8).
• SHA-256: 1f513a69287e05c689834d7a04d10de21bd9dca43565fd5a0e962383a2d58b21

See docs/MODELS.md for the benchmark behind this choice.