I run a small browser-based photo cleanup tool called MagicRemover, and since a few people here build for the same audience, I wanted to show how it's organized.
The whole thing is deliberately boring: upload a JPG, PNG or WebP, brush over what you don't want, and the AI erases it and rebuilds the background behind it. No signup, no watermark on the export, full-resolution download. Six free erases per day (three manual brush, three AI auto-detect), paid tiers only remove the daily cap.
The six one-tap modes are where most of the value is, because each mode tunes the model for a different kind of target:
- Object — general clutter: trash, wires, cars, signs
- People — passersby and photobombers; there's a dedicated unwanted people remover page tuned for travel and street shots
- Text — date stamps, captions, UI text in screenshots
- Watermark (photo) — supplier images and stock stamps
- Watermark (video) — clips that come back with an app or agency mark burned in; the video side is built for watermark-free video in a few clicks
- Logo / Sticker — brand marks on packaging and stickers on product shots
A few implementation notes that might be useful to anyone building similar tools:
- Auto-detection matters more than brush quality. Most users don't want to trace an outline; they want to tap the thing they hate. The auto mode finds the target, the user confirms.
- Large-area removal is a separate engine. Basic inpainting smears on big regions, so there's a second HD pass for those cases.
- Before/after sliders before download cut support questions dramatically — people trust a result they can drag.
Happy to answer questions about the tool or the mode design.
Top comments (0)