The Transparent PNG Guide

What an alpha channel actually is

A normal image stores three numbers per pixel: red, green, blue, each typically 0 to 255. A transparent image stores a fourth: alpha. Alpha does not describe color. It describes opacity, how much of this pixel gets drawn over whatever is behind it.

• Alpha 255 (or 1.0): fully opaque. Draw this pixel, ignore what is behind it.
• Alpha 0: fully transparent. Draw nothing. The pixel's RGB values still exist in the file, they are just not used.
• Alpha 128: half opaque. Blend this pixel's color 50/50 with the background.

That third case is why alpha is not a simple on/off mask. Consider a strand of hair thinner than one pixel. The sensor recorded a blend: some hair, some wall behind it. The honest representation is not "hair" or "wall", it is "60% hair". Alpha stores that. It is why a good background remover outputs a soft alpha matte rather than a binary cutout, and why hair looks natural in a PNG and shrink-wrapped when the alpha is rounded to on/off.

One consequence worth internalising: transparency is per-pixel, not per-image. A PNG does not have "a transparent background". It has pixels whose alpha happens to be 0. The image is still a rectangle. Nothing was deleted.

The format question, settled

The row that matters is JPEG. JPEG has no alpha channel. This is not a setting, a quality option, or a limitation of your software. The specification has no place to put opacity data. A JPEG is three channels, always, by definition.

GIF gets one note because people still reach for it: its transparency is 1-bit, so each pixel is fully visible or fully invisible with nothing between. No soft edges, so every GIF cutout has jagged edges that only look right on the exact background color you designed against. Use PNG.

Why your PNG background is black

The most common transparency question, and the answer is nearly always the same: your image went through JPEG somewhere.

When software saves a transparent image to a format with no alpha channel, it has to do something with the transparent pixels. It flattens them, compositing onto a solid color and throwing the alpha away. The color is up to the software, and the common defaults are white and black.

Black comes from a tidy reason. Fully transparent pixels often have RGB values of (0, 0, 0) underneath them. The alpha was hiding those values, so nobody cared what they were. Remove the alpha and the black that was always sitting there becomes visible. Nothing corrupted the file. The transparency was hiding a black rectangle, and you removed the transparency.

Where it happens: exporting to JPEG, often by accident from a dropdown default; uploading to a platform that re-encodes for size; clipboard and screenshot paths that drop alpha; messaging apps that compress images.

You cannot fix the JPEG. The alpha is gone and no tool recovers it, for the same reason no tool recovers detail from a downscaled image: the data is not in the file. Go back to the version that still has transparency and export to PNG or WebP. If that version is gone, you will have to remove the background again from the flattened file, and a JPEG-flattened cutout has a fringe of the flatten color baked into the subject's edge, so the second pass will be worse than the first.

The checkerboard, and why editors show it

Open a transparent PNG in any editor and the transparent regions show a grey and white checkerboard. There are no checkerboard pixels in your file. It is a convention drawn by the software, solving a real problem: transparency has to render as something, and every solid color it might pick is a color the image itself might contain. A checkerboard is unambiguous because no photograph contains a perfect grey checkerboard.

Two practical points. Never judge a cutout on the checkerboard: it is light grey, and light grey hides exactly the errors you need to see, meaning pale halos and fringing. Put the cutout on a saturated color instead. And a checkerboard in your exported image means something went wrong, because real transparency is never visible. Someone screenshotted an editor.

Making one

The background remover outputs a transparent PNG at your original resolution. Drop an image, wait, download. The model runs in your browser, so nothing is uploaded.

It uses a matting model rather than a hard segmenter, which matters specifically for the alpha channel: it produces real intermediate opacity values on wispy edges instead of rounding every pixel fully in or fully out. That is the difference between hair that looks like hair and hair that looks cut out with scissors. The model loads once, gets cached, and every run after that starts immediately. The how-to, including how to judge a cutout, is in how to remove a background.

If your source is small, upscale before you cut out rather than after: the image upscaler gives the matting model more pixels to find an edge in, and a soft edge upscaled after the fact stays soft.

Common problems

White fringe or halo. A pale outline tracing your subject means background pixels came along with it, usually because a hard mask rounded partly-background pixels up to fully-subject. Most visible on hair and shoulders. A soft alpha matte reduces it considerably. Check on a dark saturated color. The same problem inverted gives you a dark fringe when you cut out against a dark background and place the result on a light one.

Halos that appear only after resizing. Resizing an image with transparency can blend the RGB of transparent pixels into their neighbors, dragging whatever color was hiding under alpha 0 into your visible edge. This is what premultiplied alpha exists to prevent. If edges look fine at full size and fringed after a resize, this is why.

Works locally, breaks after upload. The platform re-encoded your file. Test one image before committing to a batch.

File size, honestly

PNG is lossless: every pixel comes back exactly as it went in. That property has a cost. PNG compresses well on flat color and hard edges, which is why logos, icons, UI screenshots and line art are small. It compresses badly on photographs, where every pixel differs slightly from its neighbor and there is no repetition to exploit. A 3000 pixel photo as PNG can easily be 8 MB where the JPEG is 800 KB. If an image is photographic and has no transparency, PNG is the wrong format.

WebP is the answer when the destination is a browser. It supports alpha in both lossy and lossless modes, and that is the crucial part: you can have transparency and lossy compression, which PNG does not offer. Lossy WebP with alpha often lands at 25% to 35% of the equivalent PNG at similar quality, and every current browser supports it. The caveats are outside the browser, where some desktop software and print workflows still refuse it, which is why PNG stays the safe default when you cannot verify the destination.

• Transparency, web, you control the destination: WebP.
• Transparency, destination unknown or not a browser: PNG.
• Transparency, flat color or line art: PNG, small anyway.
• No transparency, photographic: JPEG or lossy WebP.
• Shapes rather than pixels: SVG, scales infinitely for free.

Using them downstream

Presentations are the best case. A logo or cutout person on a colored slide only looks right with real transparency, and a white box around your logo on a dark slide is the classic tell. PowerPoint, Keynote and Google Slides all handle PNG alpha correctly.

Web. PNG and WebP both work everywhere. Watch out that a transparent image inherits whatever is behind it, including your dark mode background: a cutout with a subtle white halo looks fine on a white page and terrible on a dark one.

Print. Ask first. Print is CMYK, PNG is RGB, and many shops want TIFF or a flattened PDF with transparency handled in layout rather than in the image. A file that looks right on your monitor can flatten unexpectedly at the RIP.

Compositing onto other images is where halos surface. If the goal is to make a subject stand out from its own background rather than move it elsewhere, blurring the background usually looks better than cutting it out, because it keeps the original lighting and depth instead of pasting the subject into a scene it was never lit for.

Nothing you make here is uploaded

The cutout tool runs the model inside your browser tab. Your image is read into memory, processed on your own hardware, and discarded when you close the tab. There is no upload endpoint on this site.

Verify it in thirty seconds rather than trusting us: open developer tools, watch the Network tab, run the tool. The model comes down. Your image does not go up. That is the difference between a verifiable claim and a promise. More on how private this is and how we compare to the hosted alternatives.