Drop your image here or click to browse
Supports JPG, PNG, WebP. Max 2048px. All processing happens in your browser.
How to Use Image Eraser
- 1
Upload Your Image
Drop or browse for a JPG, PNG, or WebP image. The AI model loads automatically on first use and is cached for future visits. - 2
Select the Area to Erase
Use the Brush, Rectangle, Ellipse, or Lasso tool to paint over the object, person, text, or watermark you want to remove. - 3
Erase and Review
Click 'Erase Selected' and the AI fills in the area naturally using surrounding context. You can erase additional areas iteratively. - 4
Download the Result
When satisfied, click Download Result to save your edited image. All processing happens locally in your browser.
Why Use an AI Image Eraser?
The Image Eraser is a browser-based AI tool that removes unwanted objects, people, text, or watermarks from photographs using neural-network inpainting. It runs MI-GAN (Mobile Image GAN), a lightweight 28 MB model designed by Picsart AI Research for fast inpainting on mobile devices, plus an optional 207 MB LaMa (Large Mask) model for higher-quality reconstruction. Both run directly in your browser via the WebGPU API or WebAssembly fallback, with model inference handled by ONNX Runtime Web. Unlike cloud-based alternatives that upload your photos to remote servers, this tool processes everything locally on your device, making it safe to use with sensitive, private, or proprietary images.
Model selection: when to use which. MI-GAN (the default Fast model) is excellent for small-to-medium objects with clear backgrounds — power lines in landscape photos, text and watermarks, small distractions, and skin retouching. It's roughly 7× faster than LaMa and works on phone-class hardware. LaMa (the Quality model) handles larger masks, complex textures, and reflections better — choose it for removing people from busy scenes, larger objects, or anything where MI-GAN produces visible blur. Both models work on the same canvas; switching between them is one click and re-runs the same mask.
Inpainting failure modes worth knowing. Diffusion-style and GAN-based inpainting models share predictable weaknesses: very large masks (over 30% of the image) often produce hallucinated content that doesn't match the scene; semi-transparent regions (glass, water reflections, fine hair) confuse the masking step; and regular periodic textures (brick walls, fence patterns, fabric weaves) sometimes show seam artifacts at the mask boundary. Mitigations: erase in passes (small mask first, then re-mask remaining artifacts), feather the mask edge by 5-15 px, and for periodic textures try cropping the source image so the texture extends well beyond the mask.
The editing workflow supports multiple selection tools including brush, rectangle, ellipse, and lasso for precise area selection. After marking the region you want to remove, the AI analyzes the surrounding context and fills in the area with realistic content that blends naturally with the rest of the image. The tool supports iterative editing, allowing you to erase additional areas after each operation to build up complex edits step by step. Images up to 2048 pixels on the longest side are supported in JPG, PNG, and WebP formats. The AI model is downloaded once on first use and cached in your browser for instant loading on subsequent visits. Chrome and other Chromium-based browsers provide the best performance through WebGPU acceleration, while Firefox and Safari work through a WebAssembly fallback.
Pair the Image Eraser with the Image Resizer to prep a photo before erasing, the Image Compressor to optimize the result for web upload, the Image Color Extractor to pull a palette from the cleaned image, or the Background Remover for the related task of isolating subjects rather than removing them.
How It Compares
Cleanup.pictures is a popular cloud inpainting tool but uploads every image to its servers, watermarks output above HD resolution, and gates batch operations behind a paid plan. Magic Eraser in Google Photos works only inside the Google Photos app and only on Pixel phones (or Google One subscribers). Photoshop's Generative Fill is the gold standard for quality but requires a Creative Cloud subscription and uploads images to Adobe's servers. RunwayML's Inpainting handles complex scenes excellently but caps monthly usage and routes through their cloud. Inpaint.online watermarks output and limits image size. The FindUtils Image Eraser runs both MI-GAN and LaMa entirely in your browser — no upload, no signup, no watermark, no plan-tier gating, and your photos never leave your device.
The tradeoff is that browser-side inference is slower than dedicated GPU server inference for very large images, especially on phones without WebGPU support. For one-off edits and privacy-sensitive workflows, the in-browser approach is the right pick. For batch jobs of hundreds of images per day, a paid cloud service may still be more cost-effective.