Batch background removal for large product catalogs

Removing the background from a single product photo is one of the most solved problems on the internet: drop the file in, watch the subject pop free, download a transparent PNG. Removing the background from a four-hundred-SKU catalog is a different job entirely, and the difference is not the AI — it is everything around the AI. When I processed my first real catalog through the Background Remover, the cutouts themselves took an afternoon. The rework caused by an inconsistent shoot, a crashed tab full of unsaved results, and a folder of files named IMG_4club7.png took the rest of the week.

That week taught me to treat catalog work as a production line with five stations: the shoot, the batch run, the quality pass, the naming step, and the marketplace export. Each station has one purpose — to make the next station boring. A consistent shoot makes the mask predictable. A right-sized batch keeps the browser alive. A triage-style QC pass keeps you from polishing pixels nobody will see. Disciplined filenames make the upload a drag-and-drop instead of a matching puzzle. And knowing the size specs before you export means you export once.

This guide walks the line station by station, with the numbers I actually use: how many images per run on how much RAM, what a decoded photo really costs in memory, which SKUs to flag as hard cases before the batch starts, and the pixel dimensions each major marketplace wants. None of it is glamorous. All of it is the difference between a catalog that ships this week and one that ships eventually.

The single highest-leverage decision in a catalog job happens before any software opens: locking the shoot down so every frame looks like every other frame. I mean physically locking it — camera on a tripod that does not move between SKUs, product position marked with tape on the sweep, lights at fixed power, exposure set manually so the backdrop lands at the same brightness in shot one and shot four hundred. Auto-exposure is the quiet enemy here, because it re-meters for every product and hands the segmentation model a subtly different scene each time.

The reason consistency matters so much is that the model’s behavior is consistent when its input is. Feed it four hundred photos with identical framing, backdrop tone, and lighting, and it will make the same kind of mask four hundred times — which means one QC decision covers the whole run. Feed it a mixed bag of window light, changed angles, and three different backdrops, and every image becomes its own negotiation. The mask quality per image might still be good; the predictability across images is gone, and predictability is what a batch workflow actually runs on.

A jig does not need to be expensive. Mine is a paper sweep, two softboxes that never move, a strip of gaffer tape marking the product zone, and a laminated card with the camera settings written on it so a reshoot six weeks later matches exactly. The card matters more than the softboxes. Catalogs are living things — new SKUs arrive, old photos get retired — and the batch you run in September should be indistinguishable from the one you ran in July.

Segmentation models find edges where the subject and the background disagree — in tone, in color, in texture. Your job at the shoot is to manufacture that disagreement. For most catalogs a mid-gray sweep is the safest default: it separates from dark products and light products alike, it does not reflect a color cast onto the subject the way a saturated backdrop does, and it keeps contact shadows soft enough that the mask does not have to fight them. Pure white sweeps photograph beautifully but cost you separation on white and glass products exactly where you need it most.

Some SKUs are structurally hard no matter what you do: clear glassware, chrome fittings that mirror the room, white-on-white textiles, anything with fine mesh or fringe. I flag these before the batch run rather than discovering them inside it. The flagged group gets shot with extra care — a darker backdrop insert, a rim light to draw the silhouette — and gets budgeted double QC time. The capability reference at NSS Background Remover covers what the tool’s modes can recover from, but recovery is always more expensive than prevention, and at catalog scale that cost multiplies by the SKU count.

The pre-flagging habit sounds bureaucratic and takes about ten minutes: walk the product shelf with a notepad, mark anything transparent, reflective, furry, or backdrop-colored. On a typical hardgoods catalog that is five to ten percent of SKUs. Knowing the number in advance changes the schedule from a guess into a plan — ninety percent of the catalog will flow through the line untouched, and the flagged minority gets the attention it was always going to demand anyway.

The Background Remover runs entirely on your machine — the write-up at How in-browser background removal works, end to end explains the full pipeline — and that design is why a catalog costs nothing per image and nothing in upload time. But it also means your machine is the render farm, and the budget that matters is RAM. A JPEG that is 3 MB on disk is not 3 MB in a batch. Decoded for processing, a 4000×3000 photo becomes width times height times four bytes of raw pixels: roughly 48 MB. The model needs an input tensor, an output mask, and working canvases on top of that, so a single full-resolution image can hold 150 to 250 MB of live memory at its peak.

The tab does not get all of your RAM either. Browsers cap what a single tab may hold, and other tabs, the OS, and whatever else is open all take their share first. This is why the failure mode of an oversized batch is not a slow run but a dead tab — the browser kills the process rather than let it eat the machine. The first time it happens with forty unsaved cutouts queued, you learn the lesson permanently: export as you go, never let finished work accumulate only in memory.

The practical upside of on-device processing shows up in the same arithmetic. The model loads once, at the start of the session, and every subsequent image skips that cost — so image two hundred processes exactly as fast as image twenty, with no rate limit, no per-image credit, and no queue behind other customers. The constraint is honest and local: your RAM, your CPU, your session. Plan around it and the throughput is genuinely competitive with paid cloud APIs; ignore it and the browser will plan for you.

My chunk sizes come from crashing tabs so you do not have to. On a 16 GB machine processing typical 12-to-24-megapixel product photos, 20 to 30 images per run is the comfortable ceiling — the run completes, exports cleanly, and leaves headroom for the browser itself. On 8 GB, halve it: 10 to 15 images, and close everything else first. These numbers are deliberately conservative, because the cost of a crashed run is not the crash — it is re-verifying which images finished, which exported, and which silently did not.

Between chunks I export everything, verify the files landed on disk, then start the next chunk in the same session. If the machine has been running batches for an hour and feels sluggish, a fresh tab resets the tab’s memory watermark and costs one model reload — about the price of two images. The walkthrough at The image and video utility tools: resize, compress, upscale, stabilize covers the queue controls themselves; the discipline layer on top is mine, and it fits on an index card.

The rules below are that index card. They look almost insultingly simple, and each exists because I violated it once at scale and paid for the education.

Quality control on four hundred cutouts fails in a specific way: you inspect the first dozen like a jeweler, fall behind, and wave the last three hundred through blind. The fix is to make QC a triage instead of an inspection. Every image gets the same fifteen-second look and lands in one of three buckets — pass, fix, or reshoot — and nothing gets repaired during the pass itself. Repairs are a separate work session with its own rhythm. Mixing judging and fixing is how a two-hour QC pass becomes a two-day one.

The fifteen-second look is structured: view the cutout against the checkerboard, then against pure white, then zoom to 100% on the two or three places this product type typically fails — the underside where contact shadow lived, any handle or strap enclosing a hole of background, and whatever fine detail the product carries. Checkerboard reveals missing pixels; white reveals leftover dark halo; the zoom points catch the failures that thumbnails hide. Anything that survives all three is a pass, and most of a well-shot catalog will be.

The fix-rate is also your feedback loop to the shoot. When one product family keeps landing in the fix bucket for the same reason — say, every brushed-steel SKU losing its top edge into the backdrop — that is not a QC problem, it is a lighting problem, and ten minutes at the shooting table beats an hour of per-image repair. I keep a tally by failure type during triage precisely so the pattern is visible by the end of the pass instead of half-remembered a week later.

The fix bucket, sorted honestly, is smaller than it feels — usually ten to twenty percent of a consistent shoot — and the repairs cluster into a handful of shapes. Leftover backdrop in an enclosed region, like the gap inside a mug handle, is the most common and the fastest: a manual erase inside a zoomed view, done in seconds. A soft halo along one edge where the backdrop was brightest responds to a light edge cleanup. A missing chunk of the product itself, where the model mistook low-contrast subject for background, means restoring from the original — the tutorial at Choosing the right mode and refining edges in NSS Background Remover shows the restore-brush workflow that makes this a paint operation rather than a redo.

Batch the repairs by type, not by SKU order. Doing all the handle-hole erases in a row, then all the halo cleanups, keeps your hand calibrated to one motion and one zoom level, and the per-image time drops noticeably by the fifth repeat. It is the same logic as the chunked processing run applied to human attention: context switching is the expensive part, so arrange the work to switch as rarely as possible.

Set a per-image time limit for repairs and hold it — mine is three minutes. Any cutout that needs more than three minutes of handwork was mis-bucketed and belongs in reshoot, because a reshoot with corrected lighting takes ninety seconds at a jig that is still standing. The limit feels arbitrary until the day you spend twenty minutes rescuing one difficult edge and realize the reshoot would have produced a better result in a tenth of the time.

Naming is the least interesting station on the line and the one that determines whether the final upload takes an afternoon or a weekend. Marketplace bulk-upload flows, inventory spreadsheets, and your own future self all need the same thing: a filename that identifies the SKU and the view without opening the file. The scheme I settled on is lowercase throughout, hyphens as the only separator, the real SKU code first, then a view token, then a zero-padded sequence — nvs-mug-blk-front-01.png. No spaces, no underscores mixed with hyphens, no camelCase, because every system tolerates lowercase-with-hyphens and at least one system will mangle anything fancier.

The zero-padding earns its keep the first time anything sorts your files. Ten unpadded images sort as 1, 10, 2 — and a spreadsheet paste built on that order silently mismatches every SKU from the second row down. Two digits covers any sane per-SKU image count. The view token — front, back, detail, scale — matters because marketplaces slot images by position, and a name that encodes the view lets you build the position mapping in the spreadsheet instead of by eyeballing thumbnails.

Folders follow the same principle: originals, cutouts, and marketplace-sized exports live in three separate directories, and nothing ever gets saved over an original. The originals folder is read-only in spirit — it is the asset you paid a shoot day for, and every downstream file is regenerable from it. When a marketplace changes its spec or a template redesign wants tighter crops, the catalog gets re-exported from originals in an hour precisely because nobody ever "just quickly fixed" a source file in place.

Do the renaming at export time, in one sitting, against the SKU spreadsheet — never during the shoot and never during QC. Renaming while shooting slows the camera work; renaming during triage pollutes a judgment task with a clerical one. A single dedicated naming session, spreadsheet on one monitor and files on the other, takes twenty minutes for a few hundred images and produces a folder where any file can be found by anyone in seconds. The spreadsheet then becomes the manifest for the upload, and the whole catalog is auditable: every SKU row either has its expected files or visibly does not.

Every marketplace publishes pixel requirements, and reading them before the export run means you export once instead of three times. Amazon wants the main image on pure white — RGB 255,255,255, not merely white-ish — with the product filling at least 85% of the frame, and at least 1600 px on the longest side to qualify for zoom; 500 px is the hard floor, but zoom measurably helps conversion, so 1600 is the real target. The full spec, including the category quirks, is broken down at Amazon product photo requirements (2026): the complete spec, explained — worth reading in full if Amazon is your main channel.

Etsy renders listing images at a 4:3 ratio and recommends 2000 px on the shortest side, which leaves croppable margin for its zoom viewer. eBay floors at 500 px on the longest side and recommends 1600 px, same zoom logic as Amazon. Shopify stores are your own rules, but 2048×2048 square has become the de facto standard because it survives every theme’s crop behavior. The pattern across all of them: export larger than the floor, land near the recommended zoom threshold, and keep the product’s frame-fill consistent so your listing grid looks like a catalog instead of a garage sale.

I keep the specs in the same spreadsheet as the SKU list, and the export session works down the sheet: white-composite at Amazon dimensions, transparent PNG masters archived, per-platform sizes generated from those masters. The platform-by-platform walkthrough at Prepping product photos for Amazon, Etsy, and Shopify covers the composite-and-resize mechanics inside the tool. By this station the line has done its job — the shoot was consistent, the batch survived, QC was honest, the names are clean — and the export is exactly what the last station of a production line should be: boring, fast, and final.