Question 1

Is my photo uploaded to any server?

Accepted Answer

No. The photo is read and processed in your browser, inside a Web Worker. The internet is only used to download the cutout model (from Hugging Face) and the ONNX Runtime engine (from jsDelivr), once each. You can verify it in your browser's Network tab: after the model files are downloaded, no request carries your image.

Question 2

What is downloaded the first time, and how big is it?

Accepted Answer

In fast mode, 6.6 MB for the MODNet model plus about 12 MB for the inference engine. In high quality, 98.5 MB for BiRefNet lite with WebGPU (plus about 25 MB for the engine) or 191.9 MB on the CPU (plus about 12 MB). The model is stored in the site's storage (OPFS) and is not downloaded again unless you delete it. The engine is normally kept in the browser cache.

Question 3

What is the difference between fast and high quality mode?

Accepted Answer

MODNet was trained for portraits: it is great for people, quick and light, but it fails on products, animals and objects. BiRefNet lite segments objects of any kind and gives better edges and hair, but is about 15 times larger and, without WebGPU, took about 12 s per photo on a test desktop and takes longer on weak devices. For a photo of a person, start with fast; for a product, a pet or a complex scene, go with high quality.

Question 4

Why does high quality use a 512 px version instead of the 1024 px one?

Accepted Answer

Because the 1024 x 1024 one does not fit in a browser. In our tests the 1024 px ONNX graph uses about 7 GB of RAM on an ordinary CPU and fails with an out-of-memory error in WebAssembly, which is capped at 4 GB. The 512 x 512 px version (same architecture, re-exported, MIT license) uses about a quarter of that and runs. Since the mask is then refined against the photo itself, the edge difference is small on most images.

Question 5

Does it work on a phone?

Accepted Answer

Fast mode works on almost any recent phone: it is a 6.6 MB file and inference takes a few seconds. High quality is only recommended with WebGPU and spare memory; without WebGPU it uses about 2.5 GB (measured in testing) and can close the tab, so the tool warns you before downloading. On phones the working resolution is lower (mask up to 1536 px, colors up to 3072 px) to save memory.

Question 6

Which file formats and sizes are accepted?

Accepted Answer

Input: JPG, PNG, WebP, AVIF, GIF, BMP and TIFF, up to 50 MB and 40 megapixels per photo (the browser canvas limit). Output: PNG and WebP keep transparency; JPG is only used when the background is opaque (color, gradient, blurred or image), because JPG has no transparency. The ZIP carries all finished files at once.

Question 7

How do I get a white background for Mercado Livre or Amazon listings?

Accepted Answer

Pick the Mercado Livre (1200 x 1200 px) or Amazon (2000 x 2000 px) size and the tool turns on the white background, crops tight to the subject and centers it with a margin. Mercado Livre's developer documentation asks for 1200 x 1200 px (minimum 500, maximum 1920, up to 10 MB); an Amazon employee stated in the Seller Forums an RGB 255, 255, 255 background, the product filling 85% or more, and a 1000 px minimum. These rules change by category and country, so check Seller Central before uploading. We did not include Shopee sizes because we could not find an official page to cite.

Question 8

The cutout got a part wrong. How do I fix it?

Accepted Answer

Switch to Keep mode and paint over what the model removed, or to Erase and paint what is left of the background; Ctrl+Z undoes. For hair strands, raise the edge snap; for a pale haze around the subject, raise Clean up leftovers; for a light outline, use Shrink and Remove the old background color. Glass, smoke and transparent objects remain hard for any cutout model: there the brush is the way.

Question 9

Is the photo's resolution kept?

Accepted Answer

Colors are, up to 4096 px on the longest side (3072 px on phones); above that the photo is reduced to fit the browser's memory. The mask is computed at up to 2048 px (1536 px on phones) and scaled to the final size, which is enough because the model sees the photo at 512 px anyway. With a fixed output size larger than the subject, it is enlarged and the tool warns that it may look softer.

Question 10

Can I use the cutouts commercially? What about the model licenses?

Accepted Answer

The tool is free and adds no watermark. The models have licenses that allow commercial use: MODNet is Apache-2.0 and BiRefNet is MIT. We deliberately avoid non-commercial models such as Bria's RMBG. The rights to the photo stay with whoever took it; images of other people and of third parties need the usual permission.

Image Background Remover