toolgarden.xyz
中文

Image object eraser

Free online tool to paint over an unwanted object in a photo and have a local browser AI model rebuild just that area, then export JPG, PNG, or WebP

Erase settings

Repair model

Uses the MI-GAN 512 ONNX pipeline for better reconstruction. The first run downloads about 28MB of model files.

Brush size48 px

Press and drag on the image to paint over the object, covering its edge shadow too.

Edge blend12px
Output format
Supported formats
JPGPNGWebPGIFBMPSVGAVIF

The AI model loads and runs inside your browser. Images are never uploaded, and data is cleared when you close the page.

Paint the area to erase

Waiting for an image

Upload an image, then paint over the object you want removed.

Erase result

The erased image will appear here

About this tool

Paint over an object and the model rebuilds only the pixels you marked, leaving the rest of the image bit-for-bit identical. Compared with a rectangular watermark selection, a brush follows an irregular outline, so a passer-by, a cable, or clutter on the ground does not force a large block of usable background through the model as well.

The rebuilt area is the model inferring from surrounding pixels, not a recovery of what was hidden. Cover the object plus its shadow and the edge where it meets the background: missing a sliver leaves a ghost, while painting far too wide removes the neighbouring evidence the fill depends on.

How to use it

  1. Mark the area

    Set the brush size and paint along the object, including its shadow or reflection at the edges. Undo removes the last stroke if you overshoot.

  2. Erase object

    Run the erase and the model rebuilds the painted region locally. The first run downloads the model files.

  3. Check the seam and export

    Zoom into the rebuilt region to confirm texture continuity, repaint any remaining fragment and rerun, then choose an output format.

Supported range and limits

How it works
Paint over the object with a brush and the model rebuilds only the painted pixels
Repair models
MI-GAN 512 and LaMa, two ONNX models running locally in the browser
Output
JPG, PNG or WebP at the source dimensions
Works well on
Regular, continuous backgrounds where the object covers a small share of the frame
Works poorly on
Objects over faces, text or dense texture, and painted regions covering much of the frame
On failure
If neither model can run on the device the tool reports an error instead of returning an image that skipped AI repair

When you would use it

  • Removing a stranger from the background

    A pedestrian or car that wandered into a landscape or architecture shot is painted along its outline and the background is filled back in.

  • Clearing clutter from floors and walls

    Power strips, bins, or temporary signage left in a location shot come out, leaving a clean plate of the scene.

  • Taking out cables and poles

    Against a regular sky, overhead lines and posts crossing the frame usually fill in convincingly with a thin brush.

What to know before you start

  • The larger the painted region, the less neighbouring evidence the model has, and the more likely the fill turns blurry or repeats texture.
  • Faces, text, and regular geometric lines cannot be reconstructed accurately once covered; the output is plausible-looking synthesis only.
  • Erasing the same spot repeatedly accumulates blur. Restart from the source and repaint rather than processing an already-erased result again.

Related concepts

inpainting
An image-repair process that estimates the pixels of a missing region from surrounding structure and texture.
mask
A binary image marking which pixels must be rebuilt, produced here by the brush strokes.

Frequently asked questions

Does the object eraser upload my file?
No. The image is decoded, rebuilt by the AI model, and exported inside your browser, so the file never reaches a server. The default quality model downloads about 28MB on first use and is then usually cached by the browser.
How precise do the strokes need to be?
Cover the object plus its shadow, reflection, and the edge where it meets the background. Missing a sliver leaves a ghost, while painting much wider removes the neighbouring background the model needs as reference, which makes the fill softer.
Which objects come out cleanest?
Those over regular, continuous backgrounds that occupy a small share of the frame, such as cables against sky or clutter in front of a flat wall. Objects sitting over faces, text, or dense texture cannot be reconstructed accurately.
Is this the same as the watermark remover?
No. Both use the same repair models, but the watermark remover takes a rectangular selection, which suits tidy watermark blocks. The eraser paints, which suits irregular outlines like people or cables and changes fewer surrounding pixels.
How is this different from the watermark remover?
Both use the same two repair models; they differ in how you specify the region. The watermark remover drags a rectangle, which suits tidy watermark blocks. The eraser paints, which suits irregular outlines like people or cables and leaves more of the surrounding pixels untouched.
The erased spot came out blurry; why?
The painted region was too large relative to the background the model can reference. It can only infer content from around the region, so a wider area means a weaker basis for the guess. Erase smaller regions in several passes, or start with the most noticeable part.

Related tools