Background remove · Client

Product photo background remover

Drop a product photo shot on a uniform white background; the tool removes the background and gives you a transparent-background PNG. Phase 1A version uses a tolerance-based magic-wand approach (Canvas, no upload). The AI-powered version follows in Phase 3.

Last updated 2026-05-09. Math runs in your browser, no data leaves your computer.

General guidance only, not legal or professional engineering advice. Verify against the cited primary sources (IMDG, REACH, ChAFTA, RCEP, Customs Tariff Act, supplier SDS, etc.) before committing to a shipment, declaration, or contract. Sourzi assumes no liability for outcomes based on these calculators.

Phase 1A version, magic-wand approach

This tool works well on product photos with a uniform white or light background. For complex backgrounds, factory floor shots, or cargo whose colour overlaps with the background, the AI-powered version planned for Phase 3 will work better. For now, this is a fast pre-processing step for catalogue-grade product shots.

JPG, PNG, or WebP. Maximum 8 MB. Image stays in your browser; nothing is uploaded.

30

30 works for most clean white backgrounds. Higher (40 to 60) for backgrounds with shadow; lower (15 to 25) when light cargo is being stripped.

Stay in the loop

Want to know when the AI-powered version (better edges, complex backgrounds) ships? Drop your email.

The magic-wand approach

The tool reads the four corner pixels of the image (top-left, top-right, bottom-left, bottom-right), averages them to a single background colour, then iterates over every pixel in the image and asks: is this pixel within tolerance of the background colour? If yes, set the alpha to zero (transparent). If no, leave the pixel intact. The output goes onto a second Canvas and is offered as a PNG download.

This works for product photography because the photographer almost always centres the cargo and uses a uniform light backdrop, which means the four corners reliably represent the background. It fails when the cargo touches the edge of the frame (a corner samples cargo not background), when the backdrop has a gradient (one corner reads 240,240,240 and the opposite corner reads 220,220,220), or when the cargo is the same colour as the backdrop (white sacks on white).

Tolerance is the colour-distance threshold expressed in RGB space (Euclidean distance, sqrt of squared deltas). 30 means a pixel within sqrt(30^2 * 3) = 52 RGB units of the average background colour is treated as background. Camera sensor noise on a real photo is typically 5-15 RGB units, JPEG compression adds another 5-10, so 30 catches most legitimate background pixels without over-cutting into the cargo.

The hard edge between background and cargo is what gives the result a slightly pixelated look at high zoom. ML-based segmenters (rembg, MediaPipe Selfie Segmentation, ISNet) produce soft alpha masks that blend smoothly across the boundary. The Phase 3 server-side or WASM version trades bundle size or server cost for that edge quality. For most product-catalogue use cases, the Phase 1A magic-wand version is enough.

Worked example. Three drums on a white wall

The booking. A US distributor receives a product photo from a Chinese supplier showing three 200 L blue HDPE drums against a white factory wall. The distributor wants to drop the photo into a product catalogue with a clean transparent background so it composites cleanly against the catalogue layout. Original photo is JPG, 1200 by 800 pixels, white background lit by overhead fluorescent.

The work. Distributor drops the JPG into the tool. Tolerance defaults to 30. Tool samples the four corner pixels (all read as 245, 245, 240 to 250, 250, 245) and averages them to roughly 247, 247, 243. Iterates over the 1200 by 800 = 960,000 pixels and sets the alpha to 0 for any pixel within tolerance distance of 247, 247, 243. The blue drums sit far from the background colour in RGB space (about 50, 80, 180), so they survive intact. Result PNG is 1200 by 800 with transparent background and the three drums in clean cut-out form.

The edge case. Distributor notices a faint white shadow ring around each drum where the photographer used a soft light. Sliding tolerance from 30 up to 45 catches the shadow as background and removes it. The drums are clean. If tolerance had been pushed too high (60+), the tool would have started chewing into the white-tipped lid of each drum. The right answer for this photo is 45, found in two slider movements.

Frequently asked

What kind of images does this tool work best on?

Product photos shot on a uniform light background (white or near-white). The current Phase 1A version uses a tolerance-based magic-wand approach: it samples the background colour from the four corners of the image and removes pixels that match within a tolerance window. It works well on a flat-lit drum or IBC photo against a white wall, and poorly on cargo against a complex factory floor.

When does this tool fail and what should I do?

It fails on cluttered backgrounds, gradients (a backdrop with shadow), and on cargo whose colour overlaps with the background (white sacks against a white wall). The fix is either to re-shoot on a clean white surface, or to use the upcoming AI-powered version that ships in Phase 3 when the WASM bundle vs server-side trade-off is decided. For now, the cleanest workflow is to use this tool on photos already shot for catalogue use, and to budget a manual touch-up in your image editor for difficult cases.

Why a tolerance slider?

Background colour rarely sits at a single RGB value. A "white" wall photographed under tungsten lighting reads as 240, 235, 220 on RGB, with 5-10 RGB points of variation pixel-to-pixel because of camera sensor noise and JPEG compression. The tolerance slider widens or narrows the colour window the tool treats as background. Start at 30, raise to 50 if edges are too crisp, lower to 15 if clean foreground pixels are getting stripped.

Does the tool send my image anywhere?

No. The image stays in your browser memory; pixel processing runs on the local Canvas API. The output PNG is generated locally and offered as a browser download. There is no server upload at any point.

Why does the output look pixelated at the edges?

Hard-edge alpha removal is the trade-off of the magic-wand approach. The tool either treats a pixel as background (alpha 0) or as foreground (alpha 255); there is no soft transition for fringing. For higher-quality edges with proper alpha blending, you need an ML-based segmenter. That is the Phase 3 upgrade.

Related tools