Remove or replace backgrounds on a lot of images

$0.02 an image through the API, and a batch mode in the app if you'd rather not write code.

Use it when

You have volume and you want it handled, not administered. This is the default unless one of the reasons below applies to you.

Cost

$0.02/image on the API.API: $0.02 per image for background removal, $0.10 per image for the generative editing endpoints. Free API tier: 1,000 sandbox calls a month (watermarked) plus 10 free production removal calls. Enterprise needs an annual commitment and a 200,000-image minimum. The consumer app has its own tiers — check current pricing.

Good at

Products, and having two front doors. The same engine is reachable from a phone app, a web batch tool, a Shopify app and a REST API, so you can start by hand and automate later without changing tools or results.

The catch

Two price lists, one API, and it's easy to be on the wrong one. Removal is $0.02; the moment a call touches AI backgrounds, shadows or relighting it becomes a $0.10 call, and the bill is where you find out. Unused images don't roll over month to month, and the enterprise rate is fenced behind a 200,000-image annual commitment, so mid-volume users pay retail.

Wrong for

Images that can't leave your network, and any volume where $0.02 each stops being rounding — at a million images that's $20,000 to do something a GPU in the corner does free.

Free at any volume, on your own machine, with the current best open model behind it.

Use it when

The volume is large, the images are sensitive, or the per-image price of anything else has stopped making sense. Runs as a CLI, a Python library, an HTTP server or a Docker container.

Cost

Free — you pay for hardware and setup.Free. You pay for the hardware and the hour it takes to set up.

Good at

Batch as a first-class command — point it at a folder and walk away. Model choice is yours: birefnet-general, birefnet-portrait, u2net, isnet, silueta, SAM. GPU support through ONNX Runtime on both CUDA and ROCm; CPU works, slowly.

The catch

The package is MIT but the model weights are not — each checkpoint carries its own licence, independently. rembg's current default checkpoint is bria-rmbg, which needs a paid commercial agreement, so the out-of-the-box configuration is the one you may not legally ship. Switch to a BiRefNet variant, which is MIT, and check it yourself rather than trusting this sentence. Also: on CPU it's seconds per image, and when a matte comes out wrong there is nobody to email.

Wrong for

Anyone who doesn't want to run software. There is no web page here — it's a terminal, a Python environment and a licence question.

The cheapest batch button for someone who is never going to open a terminal.

Use it when

You have a few hundred images a month, no engineer, and no interest in per-call pricing.

Cost

Free tier, then $10–$30/mo.Free tier with limited background removal and watermark-free exports. Pro $10/mo ($8 yearly) with 1,000 batch exports and 600 AI credits a month. Business $30/mo ($24 yearly) with 2,000 batch exports and 3,600 credits. An API exists with separate terms.

Good at

Flat, predictable pricing for click-based batch work. A thousand exports for $10 is the best ratio here for someone who wants a UI rather than an endpoint.

The catch

The caps are hard and they're monthly. A thousand batch exports sounds generous until a re-run of a batch you already processed counts against it again, and there's no rollover. The free tier's background removal is described only as 'limited', with no published number — you find the ceiling by hitting it.

Wrong for

Volume past a couple of thousand a month, or anything you want to run unattended on a schedule.

The tool that made one-click cutouts normal. Priced for occasional use, not for a folder.

Use it when

You need a handful of very good cutouts and you want the least possible thought involved. Still excellent on hair.

Cost

Free tier is preview-resolution only.Free web tier returns preview-resolution output; full resolution costs a credit. First 50 API calls a month are free. Beyond that it's credits or a subscription — check current pricing, as the per-image rate varies sharply with volume.

Good at

Raw accuracy on the hard cases, especially hair against a busy background. Handles up to 50MP on JPG, WebP or ZIP output (10MP on PNG), and the API will take up to 500 images a minute at low resolution.

The catch

The economics are the problem, not the model. It's among the most expensive per image in this list, and the throughput limit scales down with resolution — a 50MP image drops you to 10 a minute, so your big files are also your slow ones. The free tier's preview-resolution output is unusable for anything printed or zoomed, which is a demo, not a free tier.

Wrong for

Bulk. At folder scale you are paying a premium per image for a quality difference that BiRefNet has largely closed.

Arguably the best matte here. The licence is why it isn't higher.

Use it when

Edge quality is contractual — you're delivering cutouts to a client, or the alpha channel feeds something downstream that can't tolerate a halo.

Cost

Free for non-commercial only. Commercial priced on request.Weights are free to download for non-commercial use only. Commercial use needs either a purchased self-hosted licence or the BRIA API — both priced on request, so check current pricing.

Good at

Fine edges and semi-transparent objects. Trained on 15,000+ professionally labelled images and outputs a proper 8-bit alpha matte rather than a hard mask, which is what makes it composite cleanly.

The catch

CC BY-NC 4.0. The weights are downloadable, free and openly hosted, and using them in your business is a licence violation unless you buy an agreement. It is very easy to pull this model through a wrapper, ship it, and not realise. If money changes hands anywhere near the output, you need the commercial licence.

Wrong for

Hobby projects that might turn into a business later, and any team without someone who reads licences.