AI photo editing in the browser: how it works and what to expect

The first time you use an AI tool, the runtime engine (about 14 MB, shared by all of them) and that family's model are downloaded. Both are served by QuickToolio itself, with no third-party services, and the browser usually caches them for later visits. Your image, on the other hand, is decoded and processed in the browser's memory: it isn't uploaded anywhere.

Everything runs on a single WebAssembly thread. Removing a background or detecting faces is almost instant; upscaling can take from several seconds to about a minute on a modest laptop. Each image can be up to 25 MB and 40 megapixels.

Remove the background from an image

U²-Net looks at the photo shrunk to 320 × 320 pixels and decides what belongs to the main subject. That mask is then enlarged to your photo's size, so crisp outlines (a product, a person on a plain background) come out well, but loose hair, fur or sheer fabric lose fine detail. The same mask can place a white background, a colour, another image or a blur; changing the colour afterwards is instant because the photo isn't analysed again.

• It works best with a clear subject, separated from the background and well lit.
• Shadows on the floor count as background and disappear.
• If the result is for an official document, use the ID photo tool, which also crops to the exact size.

Give a photo a white background

Change the background colour

Blur a photo's background

Make an ID or passport photo

U²-Net is trained on photos of objects, not ink strokes or flat graphics. For a dark logo on white or a scanned signature, a classic method does better: measure each pixel's luminance and make everything above a threshold computed with Otsu's method transparent. It's instant, downloads no model and keeps the stroke's original colour. Its limit is the opposite case: a white logo on a dark background would vanish.

Remove a logo's background

Remove a signature's background

When you paint over an object, MI-GAN generates plausible content from its surroundings. It doesn't know what was behind it: it imagines it. That's why it works very well over sky, wall, sand or grass and worse in front of highly detailed scenes. If the photo is larger than 1400 pixels, the fill is computed at that resolution and recomposited only inside the painted area, so the rest of the image keeps its original quality.

• Paint slightly past the outline and include the object's shadow.
• Several small areas one after another work better than one huge area at once.
• It's not for stripping watermarks from other people's photos: besides the legal problem, the fill usually shows.

Remove objects from a photo

Remove people from a photo

Real-ESRGAN always upscales 4x by rebuilding plausible edges and textures, but it works with at most 600 pixels on the longest side. A 400 × 300 px thumbnail comes out at 1600 × 1200 px looking much better than a plain stretch; a 4000 × 3000 px phone photo, however, is first reduced to 600 × 450 and the result is 2400 × 1800, smaller than the original. The tool shows you the final size before processing.

The sharpening and denoising variants use the same model and return the image at its original size: they pay off with small or heavily compressed images, and on a large photo they can make it softer. On faces and small text, the model can invent features that weren't there.

Increase an image's resolution

Fix a blurry photo

The detector analyses the photo at 320 × 240 pixels, accepts faces with 70% confidence or more and enlarges each box by 25% to cover hair and ears. Very small faces in a large group photo, faces in profile or covered faces can slip through, so check the image before publishing it. A light blur can leave someone recognisable to people who know them; pixelation with large blocks anonymises better.

Blur faces in a photo

Pixelate faces in a photo