# Remove Object from Photo > Remove people, wires and clutter from a photo: brush over them and an AI model on your graphics card fills the gap. Full-size JPEG or PNG, no upload. - URL: https://bytescope.dev/en/remove-object/ - Locale: en · Also: [ja](https://bytescope.dev/ja/remove-object.md) · [zh-Hant](https://bytescope.dev/zh-Hant/remove-object.md) - Runs: Your files never leave your browser — all processing is local. - Requires: A browser with WebGPU (tested in Chrome on desktop) on a real graphics card. It does not run on the CPU: without WebGPU, or with a software renderer such as SwiftShader, the page explains why and does not start. The first erase downloads a 208 MB model and a 26 MB engine once and prepares the model for about a minute; phones with WebGPU are untested. - Updated: 2026-09-08 ## What it does Remove Object from Photo is a magic eraser that runs in your browser. Open a photo, paint over what should not be there — a passer-by behind your friends, a power line across the sky, a bin on the beach, a stray sign — and press Erase. An AI model fills the painted area with what it thinks was behind it and blends it into the rest of the picture. You can erase several things one after another, undo and redo every step, drag a divider to compare the original with the result, and save the photo at its original size as JPEG or PNG. The photo is decoded, edited and saved inside this tab; nothing is uploaded. ### How it works The model is LaMa (Large Mask Inpainting), published by Suvorov and colleagues at Samsung Research under the Apache-2.0 licence; this page uses the big-lama weights, trained on the Places365 collection of scene photos, converted to ONNX and run on your graphics card with ONNX Runtime Web. LaMa always works on a 512 × 512 picture, so instead of shrinking the whole photo the page cuts a square around what you painted, 2.5 times the size of the painted area so the model sees enough of the surroundings, and scales that square to 512. The mask is widened by a few pixels first, because the model does badly when the mask hugs the object's edge exactly. The result is scaled back and pasted only inside the widened mask, with a 3-pixel soft edge. Every pixel outside that edge stays exactly as it was in your photo. ### What it does well, and what it does not The model guesses what was hidden; it cannot know. Simple backgrounds — sky, a wall, grass, sand, water, snow — usually come out seamless. Crowds, text and regular patterns such as tiles, fences or rows of windows often come out blurred or repeated, and a large object that covers about half the frame is rarely erased cleanly. Paint generously and include the object's shadow and reflection, or they stay behind. If one pass is not clean, undo and paint a little wider, or erase what is left in a second pass. ### What your computer needs The first erase downloads the model once, about 208 MB, plus about 26 MB for the inference engine (ONNX Runtime Web, from the jsDelivr CDN); both are then kept in this browser's cache. The model is then prepared for your graphics card, which took about a minute on a GTX 1060, and the very first erase about half a minute more; the prepared model is kept while the page stays open, and each later erase takes about three seconds. It only runs through WebGPU: on the processor a single erase took more than a minute, so there is no CPU mode, and a browser without WebGPU, or with graphics acceleration turned off, gets an explanation instead. Photos up to 8192 pixels on the long side and 40 megapixels are accepted. The saved file is re-encoded from the edited picture, so EXIF data such as the location and camera details is not carried over. ## Limits Does not upload photos, does not know what was really behind an object (it guesses, so crowds, text and regular patterns can come out blurred or repeated, and objects covering about half the frame are rarely erased cleanly), does not select objects automatically, does not generate new content from a text prompt, does not work on video, and does not carry EXIF data into the saved file. ## FAQ ### Is my photo uploaded? No. The photo is decoded and edited in this tab, and the model runs on your own graphics card through WebGPU. The only downloads are the model and the inference engine, the first time; your photo never leaves the browser and nothing about it is recorded. ### Why does the first erase take so long? Three things happen once. The 208 MB model downloads; it is then prepared for your graphics card, which took about a minute on a GTX 1060; and the very first erase takes about half a minute more. The prepared model stays ready while the page is open, so later erases take about three seconds, and on your next visit the model loads from the browser's cache instead of downloading again. ### Why does it need a graphics card? LaMa is a large network — the model file alone is about 208 MB — and every erase takes billions of calculations. On a GTX 1060 through WebGPU an erase takes about three seconds; on the processor the same erase took more than a minute. Instead of a CPU mode that would freeze the tab, the page tells you what is missing and links to My Computer Specs so you can check your browser and graphics card. ### Why does the filled area look blurry or repeated? Because the model has to invent what was hidden from what it can see around it. Sky, walls, grass and sand are easy to continue; faces in a crowd, letters and regular patterns are not, so they come out soft or copied. Paint a little wider than the object, include its shadow, erase a big object in two or three smaller passes, and undo a step whenever a result looks wrong. ### Does it keep the full resolution of my photo? Yes. Only a square around what you painted goes through the model, and only the painted area plus a soft edge of a few pixels changes; the rest of the photo is the original, pixel for pixel. You save it at the original width and height as PNG or as JPEG at quality 92. Because the file is re-encoded, EXIF data such as the location is not carried over. ### Does it work on a phone or tablet? Painting works with a finger, and the page runs where the browser has WebGPU, but it has not been tested on phones yet. Phones with little memory may fail while the model is being prepared, and the first visit downloads about 230 MB. If your browser has no WebGPU, the page says so before anything is downloaded. ## Related tools - [Remove Background](https://bytescope.dev/en/remove-background.md) — Cut the subject out and make the background transparent or a new colour. - [Restore Old Photos](https://bytescope.dev/en/photo-restore.md) — Upscale and colourise an old family photo. - [Blur Faces](https://bytescope.dev/en/blur-faces.md) — Blur faces in a photo before you share it.