Make an image background transparent

Two ways to make an image background transparent

A screenshot and a portrait need different tools. A screenshot, a logo on white or an icon on a brand color has a background that is one color, and the right cut is exact: every pixel of that color goes, every other pixel stays. A photo of a person against a wall, a bird against the sky or a bottle on a gradient has no single background color, so something has to decide what the subject is. This page offers both methods side by side and says which one it suggests. Choosing a method runs it; there is no separate start button.

The suggestion comes from the image edges. When three of the four sides are the same quiet color, the page starts Remove a flat color right away and prints that color, for example #F3F4F6 for a light app window. A title bar may spoil the fourth side without changing the answer. When the edges vary or carry texture and noise, it suggests Detect the subject and waits for you to choose, because that method downloads a model. You can switch at any time without starting over.

Remove a flat color: exact cuts for screenshots and logos

This method compares every pixel with the background color and removes the ones within the color tolerance. Everything else stays fully opaque, so a white card on a light gray page, dark text and a green button all remain solid while the page around them disappears. The same color inside a shape is removed too: the counter of a letter O, the hole in a ring logo, the space between the strokes of a signature. If that inside area belongs to the subject, turn on Only remove color connected to the image edge and it stays.

The starting tolerance is set just above the noise measured along the edges, so a clean PNG gets a tight cut and a compressed JPEG gets room for its block artifacts. Raise it when a faint halo of the background survives; lower it when a pale part of the subject starts to vanish. If the suggested color is wrong, choose Pick from image and tap the background. This method needs no model and downloads nothing.

Smooth anti-aliased edges keeps curved text and round logos from looking jagged. Only pixels right next to removed background can become partly transparent, and only when they are a genuine blend between the background and a stronger neighboring color. Those edge pixels also have the background color taken back out, so a white logo cut from blue does not keep a blue fringe. Turn the option off for a hard pixel edge.

Detect the subject: a local model for photos

For photos the page uses U²-Net small, a salient-object model that rates how likely each part of a 320 × 320 copy of the image is to be the main subject. That confidence map is interpolated to the full image size, and the Subject threshold turns it into a cut: pixels rated at or below the threshold become fully transparent, pixels above it fully opaque. Lower the threshold when parts of the subject disappear, such as a guitar body or dandelion seeds the model was unsure about; raise it when background patches remain. Edge softening fades in only the pixels just above the threshold, and at 0 every pixel is either fully opaque or fully transparent.

The model is compact on purpose. The full U²-Net checkpoint is 176 MB; this page ships the authors’ 4.39 MB small checkpoint in float precision because a quantized 1.3 MB export changed the masks materially. It is a 2020 detector, not a paid matting service, so hair against a similar wall, glass, smoke, shadows and several competing subjects are the places to inspect first. When its prediction is weak the page says so instead of presenting a confident-looking mask.

Correct what either method missed

The brushes edit the mask, never the colors. Remove paints transparency; Keep restores full opacity. Undo removes the latest complete stroke, and Reset corrections returns to the automatic mask. Strokes survive threshold and tolerance changes, so you can adjust the cut after painting. The preview always shows the whole image in its own proportions, over a checkerboard so every transparent pixel is visible.

What the download contains

The PNG has the decoded source width and height. PNG is required because JPEG has no alpha channel, and the new file does not copy EXIF, GPS, embedded thumbnails or other source metadata through the browser canvas. After encoding, the page decodes its own file again and reports how many pixels are fully transparent, fully opaque and partial. Those numbers describe the saved file, not the settings, so if they ever disagree with the preview the receipt says so.

Open the download over white and over a dark color. A pale fringe hides on white, while a dark halo hides on black. If the next step is JPEG, first decide what should replace transparency: the PNG to JPG converter lets you choose a solid background instead of letting transparent pixels become an accidental black rectangle.

Everything stays on this device

The page shell loads no model and no runtime. After you choose an image, the browser decodes it and a worker computes the mask away from the page’s interaction thread. Subject detection additionally fetches a versioned ONNX WebAssembly runtime, which is MIT-licensed, and the Apache-2.0 U²-Net small model from CreatorValet’s own origin, with a progress bar that counts real bytes. The selected image, the mask, the brush strokes and the output PNG stay in this tab, and no remote background-removal API is involved. The local-processing guide shows how to confirm that no image request leaves the browser.