Taking the background out
Background removal is the step everything else depends on: you cannot put a product on pure white, fill 85% of a frame, or add a shadow until you know which pixels are the product. Here is how it is done in the browser, and — more usefully — where it stops working.
Why a colour threshold is not enough
The obvious approach is to delete every pixel close to the background colour. It fails on almost any real product, because products contain the background colour too. A white box on white paper loses its own faces. A bottle with a white label loses the label. A shoe photographed on white loses the white sole. Threshold matching has no way to tell *the white around the product* from *the white inside it*, and the result is a cutout full of holes that you then have to repair by hand — which is where most people give up on automatic removal entirely.
What edge-connected removal does instead
The method used here starts at the frame's edges and floods inwards, taking only the region that is **connected to the edge** and close enough in colour. White inside the product is never reached, because the product's outline blocks the flood. That single change is what makes the white-label and white-sole cases work. The background colour itself is estimated from a band of pixels along the four edges, taking the median rather than the average so a stray dark object in the corner does not drag the estimate. The same code runs in the browser and in the project's test scripts, so the behaviour you get is the behaviour that was verified.
Knowing in advance whether it will work
Before removing anything, the app measures how *plain* the background is: it takes the distance of each edge sample from the estimated background colour and looks at the 90th percentile of those distances. Below a spread of 18 the background counts as plain and the removal is trustworthy. Above it — a patterned surface, a strong gradient, a shelf with objects on it — the removal is not attempted quietly and hoped for. That measurement is the honest part: it is better to tell you the background is too busy than to hand you a cutout with a chunk of table still attached to it.
Where it genuinely fails, and what to do
Three cases defeat this technique, and no amount of tuning fixes them. **Busy or patterned backgrounds**, because there is no single background colour to flood. **Product colours that match the background**, where the flood runs into the product itself — a white mug on white paper with soft light and no visible edge. And **semi-transparent edges**: hair, fur, tulle, frosted glass, mesh, where the true answer is a partial alpha value per pixel rather than in-or-out. For those, the app's paid AI cutout is the right tool and the browser one is not; that split is deliberate rather than a limitation we would rather you did not notice.
Frequently asked
How do I remove a background from a product photo for free?
If the product was shot on a plain surface, the browser tool does it locally with no limit and no watermark. Busy backgrounds need the AI cutout, which costs credits.
Why does the white inside my product survive?
Because the removal only takes the region connected to the frame edge. White that is enclosed by the product's outline is never reached.
It removed part of my product — why?
Your product likely touches the frame edge, or its colour is too close to the background at some point along the outline, so the flood ran through. Reshoot with a small gap around the product, or use the AI cutout.
Will it work on hair, fur or glass?
Not well. Those need per-pixel partial transparency rather than a hard in-or-out decision. This is a real limit of the technique, not a setting.
Does my photo get uploaded for background removal?
Not for the browser removal — that runs entirely in the tab. The AI cutout is different: it sends the image to a model to do the work. The interface says which is which before you press anything.
Can I get a transparent PNG?
Yes. Switch the background to transparent and export PNG instead of JPEG. Note that marketplaces that require white will not take a transparent file.
Listing photos that pass, made in your browser. Drop a photo in, pick the marketplace, and get a frame that matches its published rules — background off, product sized, shadow optional, and an audit that tells you which rule you are still missing.
Open Product Photo Studio