What "removing the background" actually does
An image is just a flat grid of pixels with no built-in notion of what's the subject and what's the backdrop. Removing a background means generating a mask — a per-pixel map of what to keep and what to make see-through — then discarding everything the model marks as background. The hard part was never deleting pixels; it's deciding exactly where the subject ends.
For decades that meant tracing the outline by hand with the pen tool or lasso, then feathering the edge so it didn't look cut out with scissors. Modern tools skip the tracing entirely: a neural network trained on large sets of segmented images predicts the mask in a single forward pass. GotConvert's free remove-background tool runs that model directly in your browser, so a job that used to take ten minutes in Photoshop finishes in a second or two.
Why the result must be PNG or WebP, not JPG
This is the single mistake that trips people up most. Transparency lives in an alpha channel — a fourth value stored alongside red, green, and blue for each pixel, recording how see-through that pixel is, from fully opaque to fully invisible. JPEG has no alpha channel at all. Save a cutout as a JPG and every transparent pixel gets flattened onto a solid fill — usually white or black — so you get your subject back on a colored rectangle instead of a clean cutout.
PNG solves this with a full 8-bit alpha channel: 256 levels of transparency per pixel, enough for the soft, semi-transparent edge that makes hair look natural rather than stamped-on. WebP supports the same alpha channel and usually writes it in a noticeably smaller file — Google measures lossless WebP at roughly 26% under PNG on average, and cutouts with large empty areas often shrink further. The catch is reach: WebP now works in every current browser, but some marketplaces, email clients, and older desktop apps (including parts of Microsoft Office) still won't accept it — so PNG stays the safest universal choice, and WebP is the pick when you control where the file lands and want the smaller size. GIF technically supports transparency too, but only one bit of it — a pixel is either fully solid or fully gone — which leaves jagged, aliased fringes around any curve. So for a true cutout, export PNG or WebP; only reach for JPG when you're compositing onto a solid color and don't actually need anything to show through.
How the AI handles hair and edges — and where it doesn't
The tool picks between two open-source models based on a quick on-device check for a face. Portraits go to MODNet, a matting network built specifically for people that's unusually good at wispy hair and soft edges. Everything else — products, pets, plants, logos — goes to U²-Net, a salient-object detector published in the journal Pattern Recognition in 2020. Both predict a soft-edged mask rather than a hard on/off outline, which is why the cutouts blend in instead of looking scissored.
That said, don't expect flawless results on the genuinely hard cases. Fine, flyaway hair against a busy background, transparent or reflective objects like glass and eyewear, and thin low-contrast edges all confuse segmentation — the model has to guess whether you can see through a given pixel, and it sometimes guesses wrong. Clean, high-contrast subjects on a simple backdrop come out great; a person standing in front of a cluttered shelf will need a little cleanup afterward.
What people actually use it for
Four jobs cover most of the demand. Product photos for e-commerce — Amazon, Shopify, eBay, Carousell — usually need a clean white background (Amazon requires a pure-white main image), so drop the phone shot in, pick white, and download a listing-ready JPG. Profile pictures and avatars for Slack, Discord, Notion, or LinkedIn look sharper as a transparent PNG that sits cleanly on whatever theme is showing them.
Logos scraped off a colored screenshot become reusable the moment you knock the backdrop out to transparency. And design composites — pulling a person or object out to drop onto a new background in Figma, Canva, or Keynote — skip the manual pen-tool work entirely. In every one of these, the transparent PNG is the raw material; what you put behind it is up to you.
Why running it in your browser matters
Most background removers — remove.bg, Canva, Adobe Express — upload your photo to their servers, run the model there, and send the cutout back. remove.bg's own API, for instance, works by POSTing your image file to its endpoint. Even with a company you trust, that means your face, your product shots, or a photo of an ID or document has sat on someone else's infrastructure for some stretch of time.
GotConvert loads the model into your browser and runs it locally instead. The first cutout downloads a few megabytes of model files — a small face detector to spot people, plus the segmentation model itself, on the order of 10 MB combined — cached afterward, so later runs are instant. The image data itself never leaves the tab. Same end result, zero uploaded bytes, which matters more the more sensitive the photo is.
Cleaning up the edges by hand
No automatic cutout is perfect, and the fastest fix is usually a few seconds of manual touch-up rather than hunting for a better model. Open the result in GotConvert's Paint editor to erase a stray patch the model left behind, paint back a sliver of the subject it trimmed too aggressively, or drop the transparent cutout straight onto a new solid color or image background.
For anything commercial-grade — a hero product shot, a printed piece — treat the AI pass as the 90% and finish the last 10% by hand around the tricky edges. For a Slack avatar or a marketplace listing, the raw output is almost always more than enough.