AI Magic Eraser
Paint over something you do not want in a photo — a passerby, a parked car, a trash can, a logo or watermark — and the Magic Eraser rebuilds the background behind it so it looks like the object was never there. The whole thing runs in your browser; nothing is uploaded.
Erasing an Object vs Removing the Background
These sound similar but solve opposite problems. A background remover keeps your subject and deletes everything around it, and the Smart Cutout lets you lift one object out onto a transparent canvas. The Magic Eraser does the reverse: it keeps the whole photo and makes one thing disappear, inventing plausible pixels to fill the hole. That is the job behind Google Photos Magic Eraser, Photoshop generative remove, and cleanup web tools — usually gated behind a phone upgrade, a Creative Cloud subscription, or a credit-metered service. Here it is free, with no account, and your image never leaves the page.
How AI Magic Eraser Works
Click Load model on your first visit. The browser downloads the LaMa inpainting model (about 200 MB) from the Hugging Face CDN and caches it. Drop or pick an image, then paint over the object you want gone using the brush — drag to cover it, and extend a little past its edges so the model has room to rebuild the boundary. Press Erase, and LaMa regenerates the masked region from the surrounding context. Inference runs at a working resolution capped near 1024 pixels for speed and memory, then the rebuilt patch is composited back onto your full-resolution original, so everything you did not paint stays exactly as sharp as it started — only the erased area comes from the model. On browsers with WebGPU it runs on the GPU; otherwise it falls back to WebAssembly on the CPU. When you are happy, download the result as a PNG, or send it onward to the image cropper to reframe.
The Model, the License, and Your Privacy
The tool runs LaMa (Large Mask inpainting), a resolution-robust model from the advimman project that is unusually good at filling large, irregular regions thanks to its use of fast Fourier convolutions. It runs through the ONNX export mirrored by Carve and referenced in the OpenCV model zoo. LaMa is released under the Apache 2.0 license, which permits commercial use — the bar for shipping it free on an ad-supported site, where research-only and non-commercial model licenses are off limits. Inference is handled by onnxruntime-web, Microsoft open-source runtime for ONNX models in the browser. Because the model and the compute both live on your device, your photo and the area you erase are never uploaded — not to our servers, not to the model host, not to any API. After the first load the model is cached, so the tool keeps working even offline.
Frequently Asked Questions
How is the Magic Eraser different from a background remover?+
A background remover keeps your subject and drops the entire background. The Magic Eraser is the opposite tool: you paint over one unwanted thing — a photobomber, a trash can, a stray sign, a watermark — and the model fills that area with plausible background so it looks like the object was never there. Use a background remover to isolate a subject; use the eraser to clean things out of a scene you want to keep.
How big is the download and is it cached?+
The LaMa inpainting model is about 200 MB (fp32). It downloads once from the Hugging Face CDN and your browser caches it on disk, so later visits skip the download. It is the largest model in the AI suite, which is the trade-off for running a full-quality inpainting network entirely on your device instead of on a server.
Are my images uploaded to a server?+
No. After the model finishes downloading on first use, the entire erase runs in your browser. The image and the mask you paint never leave your device and are never sent to our servers or any third-party API. Once the model is cached you can even use the tool offline.
Which model powers this and what is its license?+
It uses LaMa (Large Mask inpainting), the resolution-robust model from the advimman project, via the ONNX export mirrored by Carve and referenced by OpenCV. LaMa is released under the Apache 2.0 license, which permits commercial use. That clean license is why this can run free on an ad-supported site. LaMa is the same class of model behind popular cleanup tools.
How do I get the cleanest result?+
Paint a little beyond the edges of the object rather than tracing it tightly, so the model has room to rebuild the boundary. The eraser is strongest on small to medium objects sitting against a fairly even background — grass, sky, wall, pavement, water. Large objects, busy backgrounds, or anything with strong structure (faces, text, repeating patterns) are harder, and you may need a second pass or a touch-up in a full editor.
Why does it take a few seconds, and what is WebGPU doing?+
Inpainting is heavier than most browser AI because it regenerates pixels. On browsers with WebGPU (recent Chrome and Edge) the model runs on the GPU and is markedly faster; without it the tool falls back to WebAssembly on the CPU, which works everywhere but is slower. To keep memory and time reasonable, inference runs at a working size capped near 1024 pixels.
What formats and sizes work, and does it reduce my image quality?+
PNG, JPEG, and WebP all work. The model runs at a capped working resolution, but the filled region is then composited back onto your full-resolution original — so everything you did not paint over stays exactly as sharp as the original. Only the erased patch comes from the model. The output is a PNG.
Can I use the results commercially?+
Yes. LaMa Apache 2.0 license permits commercial use, and the edited image you export is just your own photo with one area rebuilt, so its rights are whatever they already were for your original. Nothing about this tool adds a watermark or a usage restriction.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service