Remove Background

What is Remove Background?

The thing that separates this tool from Remove.bg, Canva, and Adobe Express isn't the model — they're all using broadly similar U²-Net or BiRefNet derivatives. It's where the model runs. Theirs run on a server you pay per call or per month. This one runs the ONNX-Runtime build of @imgly/background-removal inside your browser tab via WebAssembly + SharedArrayBuffer (with WebGPU fallback on Chrome 121+ when your GPU exposes it). The first run downloads the model once (~80 MB into IndexedDB), every subsequent run is instant and offline. The practical consequence: product photos that you can't legally upload to a third-party server (NDA shots, unreleased products, client work, healthcare imagery) become processable. And the cost stops mattering — there's no $0.20-per-image meter ticking, no Pro subscription, no "watermark on free tier". You can process 200 product photos for a Shopify launch on a Saturday afternoon without thinking about quota.

Why use this tool?

Edge quality is the bar, not raw speed. The model handles hair flyaways, glass with mild transparency, jewellery with thin chains, fur on pets, and lace/translucent fabric — areas where threshold or magic-wand tools fail visibly. The output is a 32-bit PNG with soft-edge alpha, not a hard 1-bit mask, so when you composite onto a new background the transition isn't obviously cut-out. Full resolution out. No downscale, no maximum input size beyond your device RAM (a 24 MP iPhone shot at 6048×4032 processes fine on a 16 GB MacBook in 6–10 seconds; a Pixel 6a does it in 18–25 s). The model is cached in IndexedDB after first download — clear-tab won't re-fetch, only an explicit cache wipe will. Processing happens inside the tab. Nothing uploads. The /security page covers the exact data flow.

Common use cases

The dominant case is e-commerce product photography that needs to land on Amazon, Shopify, or Etsy with a clean RGB(255,255,255) backdrop — Amazon's Primary Image policy enforces pure white and rejects images that fail it. Shoot once on whatever backdrop you have, run through this tool, drop on a white canvas in any editor. Close seconds: professional headshot cleanup before LinkedIn / company-page upload (where the office wall behind you isn't the look you want); real-estate listing photos where furniture or clutter needs to go; isolating a speaker for slide deck thumbnails; preparing stock-style photos with transparent backgrounds for downstream use in Canva templates; pulling a UI element out of a screenshot to embed cleanly in design docs.

How to use Remove Background

1. 1Drop your image into the upload area (JPG, PNG, WebP, HEIC accepted; HEIC is decoded in-browser first)
2. 2On first use, the ~80 MB model downloads — typically 5–15 seconds on a normal connection. After that, no download.
3. 3Inference runs locally — 3–8 s on a recent laptop, 15–30 s on a mid-range phone. The progress bar reflects actual model progress.
4. 4Inspect the result with the checkerboard transparency view. Use the touch-up brush for any edge correction (rare; the model is good).
5. 5Download as PNG with alpha. Drop into Canva, Figma, Keynote, Photoshop, or back into /image/add-border for next steps.

Frequently asked questions

Why is the first run slow but subsequent runs fast?

The first run downloads the ONNX model (~80 MB) into IndexedDB. That's a one-time network cost. After that, the model is loaded from local storage on page load and inference starts immediately. If you clear your browser's site data, the model re-downloads on next use.

How does this compare to Remove.bg or Canva's background remover in edge quality?

For typical product shots, headshots, and well-lit subjects, the difference is invisible — all three use similar-architecture segmentation models trained on overlapping datasets. For hard cases (frizzy hair on dark backgrounds, transparent glass, motion blur edges), Remove.bg sometimes wins because they layer multiple specialised models server-side. Canva is broadly comparable to this tool. For batch work the cost difference flips the value: Remove.bg charges $0.20 per full-res image, Canva needs Pro, this one is free and unlimited.

My GPU is supported by WebGPU. Will the tool use it?

Yes on Chrome 121+ and Edge 121+ when your GPU exposes the f16/i64 features the model needs. Safari and Firefox currently fall back to the WebAssembly multi-threaded build, which is 2–3× slower but works on every modern device. The tool picks the fastest available path automatically — you don't need to configure anything.

It cut into my subject's hair / fur. How do I fix it?

Use the touch-up brush in the editor after the initial pass. Paint over the missing-hair region with a soft brush at ~30% strength to gently restore alpha. The brush composites on top of the model output so you don't lose the model's other good edges. For really difficult subjects (fluffy white pets on white backgrounds), the right move is often two passes: one with the AI to get the bulk of the cutout, then manual touchup of the contested edges.

Can I batch-process 50 photos for a product launch?

Not in a single click — the tool processes one image at a time to keep the UI responsive and avoid OOM on phones. The model stays loaded in memory between images, so each subsequent run is faster than the first by ~30% (no re-initialisation cost). For true batch workflow, the practical pattern is to leave the tab open, drop each image in, download, drop next.

Why does the output sometimes have a faint coloured fringe around the edges?

That's spill from the original background — pixels right at the edge contain a blend of subject + background colour. The model produces an alpha mask but doesn't un-blend the foreground. For product shots on a coloured backdrop, the fix is either (a) shoot on a neutral grey or (b) after removal, run the result through a colour-decontamination pass in Photoshop/Affinity. We don't auto-decontaminate because the right colour to "restore" depends on the new background you're compositing onto.

Will Amazon accept the output as a Primary Image?

Not directly — Amazon needs RGB(255,255,255) background, not transparency. Download the transparent PNG here, then open it in any editor (Canva, Figma, Preview, Photoshop) and place it on a 2000×2000 white canvas. Save as JPG at high quality. Amazon's policy requires at least 1000 px on the longest side for zoom; 2000 px is the safer target.

Which background-removal model, and where does it run?

We use the BRIA RMBG-1.4 model (transformers-compatible weights from the briaai/RMBG-1.4 release on Hugging Face), running entirely in your browser via ONNX Runtime Web with the WASM backend. Pre-trained on a proprietary corpus of ~12,000 image–mask pairs by BRIA AI; we did not retrain. Realistic accuracy ceiling: portraits and product photography against a contrasting backdrop produce broadcast-quality cutouts; the model excels at hair and translucent edges. Known failure modes: subjects whose colour matches the background (white shirt on white wall), motion-blurred edges, transparent glass, fine wire mesh, and intentionally tricky cases like chain-link fences. For each case, we expose the mask preview before download so you can verify before committing.

1. 1Shoot product photography against a slightly off-white or pale grey backdrop, not pure white. The AI distinguishes "product" from "background" more reliably when there's contrast — products on white-on-white frequently lose edge detail.
2. 2After removal, zoom into the hair / edge regions at 200% before downloading. The model is good but not perfect — a 5-second visual check catches the rare cut-into-hair case while you can still touch up.
3. 3If you're removing backgrounds from 20+ images for a launch, keep the tab open between images. The model stays warm in WASM memory and subsequent inferences are 20–30% faster than a cold start.

How does it compare?

Remove.bg gives you a 0.25 MP preview free; full resolution costs $0.20 per image or $9/month minimum. Canva's background remover requires a Pro subscription ($12.99/month). Adobe Express requires a Creative Cloud account. PhotoRoom mobile is freemium with a per-image cap. This tool runs the same class of segmentation model in your browser at full resolution, with no quota, no fee, and no upload — the trade-off is a one-time ~80 MB model download.

Related guides

Editorial walk-throughs that go deeper on the workflow most people use this tool for.

Related tools

More free tools. No signup, no watermarks.