There is no “background removal” algorithm — there are four, and they fail in different ways. Picking the right one first saves you the frustrating version of this task: clicking Undo on a tool that was never going to work on your photo.

1. Color-range selection (“erase every pixel like this one”)

The simplest idea: pick the background color, delete every pixel within a distance of it. The distance knob — tolerance — is the whole game: too low and JPEG noise leaves a confetti of leftover pixels; too high and the subject’s matching colors get eaten.

Wins on: white and grey studio backdrops, green screens, scanned documents and signatures, flat-color clipart. Fails on: gradients and vignettes (the far corner is a different color), textured backdrops, and any subject that shares the background color.

This site’s “Same color everywhere” mode is exactly this — a global color-key erase.

2. Flood fill (“erase the region connected to this point”)

Same color test, but only pixels connected to the seed get removed — think paint bucket, not magic wand. The connection rule is what saves you when the subject contains the background color: a white mug on a white backdrop keeps its white body because the mug’s whites never touch the image border.

Wins on: subjects containing background-adjacent colors, product photos with shadows, any shot where the backdrop is connected around the edges. Fails on: subjects that touch the frame with the same color (the flood leaks inside through the bridge), and backgrounds with internal islands you can’t reach from the edge.

This is the tool’s default “Connected background” mode, auto-seeded from the dominant border color — which is why most uploads remove themselves before you touch anything.

3. AI salient-object matting (“find the thing, not the color”)

A neural network trained on thousands of segmented photos estimates, per pixel, the probability that it belongs to the main subject. The output is a soft matte — fractional alpha at the edges — which is why AI cutouts feather hair and fabric better than a hard color cut.

Wins on: cluttered rooms, outdoor backgrounds, textured backdrops, pets and people on non-uniform scenes. Fails on: tiny fine detail (wispy hair, fur, glass edges at the model’s working resolution), multi-subject scenes where “the subject” is ambiguous, and product shots needing pixel-exact edges.

Here it runs as the optional AI mode: a 4.6 MB U²-Netp model in WebAssembly, downloaded once and executed on your device. It estimates at 320×320 and upscales — great for “person on a messy sofa,” less surgical than a color cut on a clean studio shot.

4. Manual masking (“draw the line yourself”)

Brush, pen, lasso — the oldest technique and still the reference standard. Nothing decides what’s subject better than you; nothing is slower either.

Wins on: jewellery, glassware, fur on same-tone backgrounds, hero images worth twenty minutes. Fails on: deadlines.

Pick by background, not by hype

Your background	First choice	Backup
White / seamless / green screen	Color range or flood fill	—
Same color as the subject	Flood fill (connected only)	Manual
Busy room or outdoors	AI matting	Manual
Soft gradients, even light	Color range, low tolerance	AI
Hair/fur against contrast	AI or color + feather	Manual

The honest rule: the more uniform the backdrop, the dumber the algorithm you need. If you’re still shooting the photos, the cheapest “AI” is a plain wall and even light — then color mode finishes the job in milliseconds. If the photo already exists with a chaotic backdrop, that’s when the ~17 MB model+runtime download earns its keep.

Refining the result: product-photo backgrounds covers marketplace white fills, and passport background rules covers the regulation angle. For a different kind of cleanup — the metadata your photo leaks — our sister tool ExifView reads and strips it on-device.

Frequently asked questions

Is AI background removal always better than color selection?

No. On a clean, fairly uniform background (studio white, green screen, seamless paper) a color-range cut is faster, more predictable and often cleaner at the edges — the model adds nothing. AI earns its keep on busy or uneven backgrounds where 'the background color' isn't a single thing.

Why do automatic removers fail on white products on white backgrounds?

Because the background color genuinely appears inside the subject. A color-based tool can't tell the product's white highlights from the wall behind it, and even AI models sometimes merge bright subject edges into a bright backdrop. The fix is contrast: shoot on grey or a complementary color, or flood-fill only the connected outer region instead of all same-colored pixels.

What is 'spill' and how do I remove it?

Spill is background color reflected onto the subject's edges — classic on green screens, where hair and shoulders pick up a green fringe. Removing the background removes the source but not the already-bounced color in edge pixels. Feathering softens it; dedicated editors desaturate the fringe hue. This site currently softens via feather rather than decontamination.

When is hand-editing still worth it?

For paid work with hard requirements: jewellery with internal gaps, glassware, fur against a same-toned backdrop, or anything where a single missed corner costs a reshoot. Automatic tools get you 90–99% there in seconds; the last percent on a hero image is still hand work with a tablet and a mask.