MagicRemover

Problem

One-click AI background removers are a black box: when they clip an edge wrong there is no way to correct just that part, and they require uploading the image to someone else's server. For images with flat backgrounds — logos, product shots, screenshots — what is actually needed is precise, repeatable control that runs locally.

Approach

A desktop tool built on classic computer vision instead of a model: click a color, flood-fill selects the contiguous region within an adjustable tolerance, and the result is written to the alpha channel. Each pick is a non-destructive layer in a history panel — adjust its tolerance or delete it individually at any time. The original and the checkerboard transparency preview share one synced zoom/pan view, and export offers resizing with aspect-ratio lock.

My role

Sole builder — product design, image-processing logic, desktop UI, and bilingual interface.

Result

A working tool I use myself, open-sourced on GitHub.

Tech stack

• Python
• OpenCV
• Tkinter
• Bilingual (EN/中文)