AI Background Remover docs
How the Toolz.dev AI Background Remover works, what it runs on your device, and where to find every setting, shortcut and fix.
The AI Background Remover cuts the subject out of a photo client-side and gives it back as a transparent PNG, WebP or JPG at the photo's full resolution. There is no upload step. The AI model downloads once, runs on your graphics card or processor, and the photo never leaves the device.
These docs are the reference for it: what each control does, which model runs on which hardware, what the numbers in the code are, and what to try when a cutout or a download goes wrong. If you only want to remove one background, the Quickstart takes about a minute.
What does the background remover do?
It takes a photo of a person, pet, product or object and returns the same photo with everything behind the subject made transparent. After that you can put the subject on a colour, on another photo or on a blurred copy of the original, resize it, crop to the subject, upscale small photos with AI, and save one file or a ZIP of many.
Three things set it apart from the usual cloud remover:
- The cutout is written at the size of your photo. A 4000 pixel product shot stays 4000 pixels wide.
- Nothing is uploaded. The model runs client-side, in a Web Worker inside the tab, on WebGPU when your browser has it and on WebAssembly when it does not.
- The edges get extra work after the model runs: stray background specks are removed and the colour of semi-transparent hair is recovered, so there is no dark fringe on a new background.
How does it work, in one paragraph?
The page decodes the photo, a worker thread scales a copy to the model's input size and runs a segmentation network that scores every pixel from "background" to "subject". On a capable GPU that network is BiRefNet lite, run twice (once on the whole frame, once zoomed in on the subject). Elsewhere it is IS-Net. The score map is cleaned, scaled back up to the photo's real size and written into the alpha channel. Models and hardware has the full selection rules and Edge quality covers the clean-up.
Where do I find what?
| If you want to | Read |
|---|---|
| Remove your first background | Quickstart |
| Learn the preview modes and every key | Compare view and keyboard shortcuts |
| Put the subject on a colour, photo or blur | Backgrounds |
| Change the output size, crop or upscale | Resize, crop and AI upscale |
| Process a folder of product shots | Batch processing and ZIP |
| Check which files go in and come out | File formats |
| Know what runs on your hardware | Models and hardware |
| Understand the speck and halo fixes | Edge quality |
| See exactly what leaves your device | Privacy |
| Fix a slow, stuck or failed run | Troubleshooting |
| Remove backgrounds from Claude, Codex, Gemini CLI or Cursor | MCP server |
What does it cost, and can I use the results commercially?
It is free, no signup, with the same daily usage limit as other Toolz.dev tools for guests. The segmentation models are MIT (BiRefNet) and Apache-2.0 (IS-Net), the upscaler is BSD-3-Clause, and Toolz.dev adds no restriction of its own to the files you make. You still need the rights to the photo you started with.
For the longer story of why cutouts go wrong at the edges and how the pipeline fixes it, the guide How AI background removal works in the browser walks through it with worked examples.