Background Remover — Remove the Background from Any Photo
Remove the background from any photo right in your browser — no upload, no sign-up, no watermark. The page loads the open ormbg model (Apache-2.0) once and then cuts out people, products, pets and logos locally, giving you a transparent PNG in about a second on a recent browser with WebGPU fp16 support. Otherwise it quietly falls back to the CPU, and photos up to 4096 px on the long side are supported.
Private by design. Your photo is processed entirely on your device and is never uploaded to any server. The only things downloaded are the ormbg AI model from Hugging Face and the inference library from the jsDelivr CDN; both are plain files, fetched once and cached by your browser, and neither contains any of your data. The page’s anonymous usage analytics never include your image or its file name.
How to remove a background
- Add a photo. Drop a PNG, JPG, WebP, AVIF, BMP or GIF (first frame) onto the page, paste it with Ctrl/⌘ + V, or press Choose image.
- Wait once on the first run. The tool downloads the ormbg model — about 88 MB when your browser has WebGPU with fp16 support, or 44 MB on the CPU — plus the inference library, each a one-time download. Your browser caches them, so later photos start instantly.
- Compare and pick a background. Drag the slider to check the cutout against the original. Keep the checkerboard for a transparent background, or choose white, black or any color you like.
- Download the result. Save a transparent PNG, or a PNG/JPG on the color you picked. Photos larger than 4096 px on the long side are scaled down to fit.
Use Cases
Frequently Asked Questions
Is it really free, and is my photo really never uploaded?
Yes to both. There is no account, no watermark and no payment, and the image never leaves your device: the cut-out runs entirely in your browser. The tool itself only downloads the AI model (from Hugging Face) and the inference library (from jsDelivr); like the rest of orangebot.ai, the page sends anonymous usage analytics, which never include your image or its file name.
Why is the first run slower than the ones after?
The first photo has to wait for the model to download: about 88 MB with WebGPU fp16 support or 44 MB on the CPU, plus the inference library. Your browser caches all of it, so the second photo on the same browser starts right away.
Why is it slow on my computer or phone?
It depends on how your browser runs the model. With WebGPU and fp16 shader support (recent Chrome and Edge on most computers, other browsers where available) the cut-out usually takes under a second. Without either, the model runs on the processor through WebAssembly and one photo takes roughly 5–15 seconds. If the GPU path fails for any reason, the tool automatically retries the photo on the CPU. The table below shows what runs where.
Which images work best, and where does it struggle?
A single clear main subject works best: people, products, pets and cars. Honest weak spots: fine hair against a busy background, glass or other transparent objects, subjects that blend into the background in color or tone, and several overlapping subjects that should all be kept.
Can I use the result commercially?
How you use the output image is your call; this tool makes no claim on it. For context, the ormbg model itself is released under the permissive Apache-2.0 license. Your product photos or artwork still need the usual rights you would need for any image.
Does it keep the original resolution?
Yes, up to 4096 pixels on the long side. Larger photos are scaled down so the long side is 4096 px before processing, and everything you download comes from that working size.
GPU vs CPU: what runs on your device
This is what the tool picks automatically when you add a photo. It checks your browser for WebGPU, uses the graphics card if WebGPU with fp16 shaders is available, and falls back to the processor otherwise. If the GPU path fails, the photo is automatically rerun on the CPU.
| Your browser | Runs on | Model download | Precision | Typical time per photo |
|---|---|---|---|---|
| WebGPU with fp16 shaders | Graphics card (WebGPU) | 88 MB | fp16 | Under 1 s |
| No WebGPU, or no fp16 | Processor (WebAssembly) | 44 MB | uint8 | About 5–15 s |
- Times were measured on an Apple Silicon laptop with a roughly 1 MP photo; phones and older laptops are slower.
- The model and the inference library download once and are cached by your browser, so only the very first photo pays the download.
- Images larger than 4096 px on the long side are scaled down before processing; everything you download comes from that working size.
Related tools
More free, in-browser tools from the Image Tools set — every one runs locally, with no sign-up and no upload.