baEraser

A GNOME image object eraser built with GTK4 + libadwaita.
Draw a mask over an unwanted object and erase it using classical inpainting or AI (LaMa).

Screenshots

img1	img2

img3	img4

img5	img6

img7	img8

img9	img10

img11

img12	img13

img14

Installation

AppImage (tested only on Solus linux)

chmod +x ./baEraser.AppImage
./baEraser.AppImage

Flatpak (recommended)

Install:

flatpak install --user -y --bundle ./baEraser.flatpak

Run:

flatpak run si.generacija.baEraser

Uninstall:

flatpak uninstall --user si.generacija.baEraser

Build from source

The build script checks the required system packages, downloads ONNX Runtime and the LaMa model when missing, then builds the executable:

bash build.sh
./build/baEraser

Manual build is also available if the dependencies are already installed:

make -j$(nproc)

Required system dependencies: GTK4, libadwaita, OpenCV 4, pkg-config, g++, make.

Inpainting methods

Method	Best for
TELEA	Quick removal, small areas
Navier-Stokes	Smooth backgrounds
Multi-scale	Large areas
PatchMatch	Backgrounds and textures
LaMa AI	Best quality (requires ONNX model)

LaMa model

Flatpak and AppImage builds bundle every .onnx file found in models/. For a source build, build.sh downloads a model automatically when none is present. Manual download is also available:

bash models/download_lama.sh

At runtime baEraser searches for LaMa models in this order:

1. BAERASER_MODEL_PATH
2. ../models relative to the binary
3. ./models relative to the current working directory
4. /usr/share/baEraser/models
5. $XDG_DATA_HOME/baEraser/models

Supported bundled model names include:

• inpainting_lama_2025jan.onnx
• lama_carve_fp32.onnx
• lama_fp32.onnx
• big-lama.onnx

When LaMa is selected but fails at inference time, baEraser falls back to PatchMatch and reports that fallback in the status bar.

License

baEraser source code is licensed under the MIT License. See LICENSE.

Release bundles may include third-party components with their own licenses, including GTK, libadwaita, OpenCV, ONNX Runtime, and downloaded LaMa ONNX models. Check the upstream projects and bundled license files before redistributing modified bundles.