Why can’t it erase a whole person — only small things?
Because this tool uses a classical inpainting algorithm (OpenCV’s Telea / Navier-Stokes), not a generative AI model. It works inward from the edges of what you painted, filling the gap using the known pixels around it — essentially extending the surrounding texture rather than understanding the scene and repainting it. So the bigger the area and the busier the background, the mushier the result. Erasing a whole person would require a model to genuinely imagine the scene hidden behind them, and that’s something only generative AI can do. Practical tip: several small passes usually beat one big one by a wide margin.
Why not use an AI model like the background-removal tool does?
We looked into it properly. The AI inpainting models that can realistically run in a browser are LaMa (ONNX around 198MB — not realistic to ship) and MI-GAN (around 26.7MB, which would be fine on size), but the repository publishing MI-GAN’s weights declares no license, and those weights were trained on a dataset whose terms state it is for non-commercial research and education only. Whether weights trained on non-commercial data may be used commercially is legally unsettled, and we don’t think that uncertainty is ours to hand to you. OpenCV’s classical algorithm is Apache-2.0 and, being pure algorithm, has no training data for that question to apply to. It’s also far lighter: the whole engine is about 3.5MB compressed, over 7x smaller than MI-GAN, so the tool is usable immediately instead of after a large model download.
How should I set “fill radius”?
It controls how wide a ring of surrounding pixels the algorithm reads when deciding what to fill in. The default suits most cases. If the patched area doesn’t blend and you can see a hard seam, nudge it up. If unrelated content from further away is bleeding in (erasing something up against a tree and the tree’s color seeps across, say), or it’s getting slow, bring it down.
Is my photo uploaded to a server?
No. Once the fill engine (OpenCV.js) has downloaded into your browser, all the computation happens locally and the photo never leaves your device. Your browser caches the engine file, so later uses don’t re-download it and it keeps working offline.