How to remove a background from an image

Under a minute, no account, and the photo never leaves your computer. Here is the full walkthrough, then a plain explanation of what the AI is doing and where it can trip.

The six steps

1. Open DropBG. There is nothing to sign up for.
2. Add your image. Click Choose images, drag a file anywhere onto the page, or paste with Ctrl+V (Cmd+V on a Mac). JPG, PNG and WebP all work. Drop several files at once if you have a batch.
3. Let the model load the first time. Your browser downloads the AI model once, about 40 to 90 MB depending on your device. A progress bar shows the download. Every later image skips this step.
4. Compare. Drag the slider left and right across the picture to check the cutout against the original.
5. Choose a background. Transparent is the default. Click white, black, a preset colour, or the rainbow swatch for any custom colour. This is handy for product photos that need a plain white background for a marketplace listing.
6. Download or copy. Download PNG saves the full-resolution file. Copy puts the PNG on your clipboard so you can paste it into Slides, Figma, Keynote or a chat.

What the AI is actually doing

The model behind DropBG is a salient object segmentation network from the IS-Net family, trained on the DIS5K dataset of finely annotated images. Given a photo, it predicts an alpha matte: for every pixel, how much of it belongs to the main subject. Solid subject pixels get full opacity, background pixels get zero, and edge pixels such as hair or fabric get values in between so they blend cleanly onto a new background.

That matte is then multiplied into your original image to produce a PNG with a real transparency channel. Nothing is painted or regenerated, which is why the output is faithful to the source at full resolution.

Why it can run in a browser at all

Two pieces of web technology make this possible without a server. WebAssembly lets the ONNX inference runtime run at near-native speed inside the page, and WebGPU, where your browser supports it, hands the heavy matrix math to your graphics card. On a recent laptop with WebGPU, most photos finish in one to three seconds. Without WebGPU the tool falls back to a smaller quantised model on the CPU, which is slower but still works everywhere WebAssembly does.

The practical consequence is the privacy guarantee: there is no upload step in the pipeline. You can confirm this yourself by opening your browser's developer tools, switching to the Network tab and processing an image. You will see the model files download once and nothing else.

Getting the best result

• One clear subject. The model looks for the most salient object. A single person or product on a reasonably distinct background gives the cleanest cutout.
• Resolution helps edges. A 2000 px photo gives the matte more to work with than a 400 px thumbnail. There is no upload cap, so use the original.
• Contrast at the edge. Dark hair against a dark wall is hard for any tool. If you can reshoot, put the subject against a plain, contrasting background.
• Check at 100%. Use the slider on tricky regions such as hair, glasses and fingers before you export.

Where automatic removal struggles

Being straight about this saves you time. Fine hair over a busy background, semi-transparent objects such as glass or veils, thin structures such as wires and fences, and subjects that fill the entire frame with no visible background are the failure modes for every automatic tool, DropBG and remove.bg included. In those cases the fastest fix is to export the PNG and clean the edge in any editor with a layer mask.

Common uses

• Product photos on white for Amazon, Etsy, eBay or Shopify listings
• Profile pictures and headshots on a brand colour
• Cutouts for presentations, decks and thumbnails
• Stickers, memes and social posts
• Logos scanned from paper that need a transparent background