A place for the workflow we relied on.
We built remover.dev because we used remove.bg ourselves. After 1 December 2026 that standalone site and self-serve API stop. The tool below is free, and photos stay on this device.
Run history—
Run each model on the same image to compare. Timings reflect this device.
| Model | Backend | Load / initialize | Inference | Total time | Result |
|---|---|---|---|---|---|
| BiRefNet Lite512 × 512 | — | — | — | — | Not run |
| BiRefNet Lite Lightweight512 × 512 | — | — | — | — | Not run |
| BiRefNet Lite1024 × 1024 | — | — | — | — | Not run |
| BEN2 Base1024 × 1024 | — | — | — | — | Not run |
| U²-NetP320 × 320 | — | — | — | — | Not run |
Total time includes startup, model loading, preprocessing, inference, and image export. Previews are limited to 1600 px on the longest edge. PNG exports are capped at 32 megapixels and 8192 px per edge; larger images are scaled proportionally. Edge adjustments reuse the original mask without another inference and are excluded from run timings. Hair refinement and color decontamination are not included.
The original stays next to the cutout.
We used remove.bg. After 1 December 2026 the standalone site and self-serve API stop. Drag the divider for what we kept, and what we needed a place for.
1 Dec 2026
The standalone site and self-serve API stop
On device
The free tool does not upload the photo
Before
Here
Drag the divider, or focus it and use the arrow keys
Your photos never leave this device.
The model downloads once, runs here, and the PNG stays here.
- 1
The model comes to you
On the first run a segmentation model downloads once, from 4.6 MB up, and the browser caches it. After that it can load offline.
- 2
Inference runs on your hardware
BiRefNet uses your GPU through WebGPU, with a WASM fallback. The photo is decoded and composited in this tab. No server sees it.
- 3
Export at full resolution
Refine the edges without running the model again, then download a transparent PNG at up to 32 megapixels. No watermark, no credits, no account.
02 — Models
Five open models, downloaded on demand.
Only the model you select is downloaded, then cached for reuse. Sizes are the one-time download; every backend runs in your browser.
| Option | Model | Input size | Download | Backend |
|---|---|---|---|---|
| Standard · 512 | BiRefNet Lite | 512 × 512 | 98.5 MB | WebGPU FP16 |
| Lightweight · 512 | BiRefNet Lite Lightweight | 512 × 512 | 75.2 MB | WebGPU FP16 |
| Detail · 1024 | BiRefNet Lite | 1024 × 1024 | 114.8 MB | WebGPU FP16 |
| Compare · 1024 | BEN2 Base | 1024 × 1024 | 219.1 MB | WebGPU FP16 |
| Compatibility · 320 | U²-NetP | 320 × 320 | 4.6 MB | WASM |
Browser models: BiRefNet and BEN2 are MIT, U²-Net is Apache-2.0. The paid API runs only U²-NetP (Apache-2.0). Bria RMBG is not shipped.
03 — FAQ
Honest answers.
Is it really free and unlimited?
The tool on this page is free and unlimited. It runs on your hardware, with no watermark and no account. A new API key includes 5 test credits. Purchased credit packs never expire. The API plan is $20 a month for 1,000 images, or $192 a year, and those monthly credits reset. Prices are on the pricing page.
Why does it download a model the first time?
Instead of uploading your photo to a model, the model comes to you. The first run fetches the network weights once (4.6–219 MB depending on the model you pick), your browser caches them, and subsequent runs start instantly — even offline. That download is the only thing that ever moves over the network.
What happens to my images?
On this page, photos are decoded and exported in this tab. Nothing is uploaded. The paid API is a different path: you send an image to POST /v1.0/removebg, it is segmented for that request, and it is not written to storage. Close the tab and the browser copy is gone.
How does the quality compare to remove.bg?
In the browser, Standard and Detail are BiRefNet, which holds up on hair and fine edges. The paid API currently runs U²-NetP, a smaller Apache-2.0 model, so treat it as a compatibility matte rather than a match for remove.bg. Results vary by image.