AI Object Remover

AI Object Remover - Brush-Based LaMa Inpainting to Remove Unwanted Objects from Photos

Remove unwanted objects, people, text, watermarks and distractions from photos using the LaMa deep-learning inpainting model. Load an image, paint a single round brush over what you want gone, adjust the brush size and the Edge Feather slider, then click Remove Objects - the AI reconstructs the painted area from the surrounding pixels. Everything runs in your browser with WebAssembly; after the page loads your photos never leave your device, so it works offline and stays completely private. Ideal for erasing watermarks, photobombers, power lines, signs, blemishes and other distractions. Compare the result with an interactive before/after slider, undo or clear your mask at any time, and download the cleaned image as PNG, JPG or WebP. Free, no watermark and no registration.

How do I remove an object with this tool?

Load a photo (upload a file or paste an image URL), then use the round brush to paint over the object you want to remove. Adjust the Brush Size slider to match the object, and use the Edge Feather slider to control how softly the repaired area blends into the surroundings. Click Remove Objects and the LaMa AI model fills the painted region using the surrounding texture. Use Undo, Redo or Clear Mask while painting, then download the result. There is one brush - paint exactly what you want gone.

What does the Edge Feather slider do?

Edge Feather controls how the inpainted patch blends into the original photo. At 0 the swap is a hard binary edge, which can leave a visible seam. Raising it dilates the mask a few pixels and applies a soft alpha falloff across the boundary, so the repaired area fades smoothly into its surroundings - the most important control for hiding the patch on watermarks, skies and skin. A value of 4-8 px works well for most photos; increase it for large fills, lower it for fine detail near sharp edges.

Which AI model powers it, and is it really private?

It uses LaMa (Large Mask inpainting), a ~208MB ONNX model run with onnxruntime-web (WebAssembly) directly in your browser. The model downloads once, then all processing happens on your own device - your images are never uploaded to any server and nothing is stored. After the first load you can even use the tool offline. This makes it safe for confidential product shots, real-estate photos, ID documents and client work.

What image formats are supported?

You can load JPEG, PNG, WebP, GIF and BMP, plus HEIC, HEIF, AVIF and TIFF where your browser can decode them - and you can drag-and-drop those files even when the operating system reports a blank MIME type. The result can be downloaded as PNG (lossless), JPG (smaller, white background) or WebP (best web compression). Output keeps the full resolution of your original image.

Does the 512px AI model limit how large my image can be?

The LaMa model runs its inference at 512x512, but the tool no longer stretches your photo into that square. The image and mask are letterboxed (padded to a square, keeping the true aspect ratio), inpainted, then the patch is cropped back and scaled up and composited onto your full-resolution original - so the output stays at the original pixel dimensions. Very large images simply take a little longer; for the sharpest fill, paint the smallest area that fully covers the object.

How do I get the cleanest possible result?

Paint slightly beyond the object's edges so no fringe is left behind, keep the masked area as small as the object allows, and let Edge Feather (4-8 px) soften the seam. Objects sitting against simple or repeating backgrounds (sky, grass, walls, water) reconstruct best. For complex scenes, remove objects in a few smaller strokes rather than one huge fill, use Undo to refine, and re-run if needed. For pixel-perfect commercial work you can export the PNG and finish in Photoshop, GIMP or Photopea.

Is there a limit on how many images I can process?

No. The tool processes one image at a time (there is no batch queue), but you can clean as many photos as you like for free - reset, load the next image and repeat. There is no watermark, no signup and no usage cap. Because all computation happens locally, throughput depends only on your device's CPU and available memory.

Does it work on mobile and touchscreens?

Yes. The brush, before/after comparison slider and all controls support touch, so you can paint over objects with your finger on a phone or tablet. The ~208MB model download and on-device inference are heavier on mobile, so a recent device with ample memory will be noticeably faster, but any modern browser with WebAssembly support will run it.