How the AI Background Remover Works
DocDox Background Remover uses a machine learning segmentation model compiled to WebAssembly and ONNX Runtime to automatically separate the subject of a photo from its background, producing a transparent PNG — processed entirely on your device. The AI model (from @imgly/background-removal) is downloaded once (~50 MB) and cached in your browser for all subsequent uses, enabling offline operation after the first run.
The model uses a deep learning architecture trained to identify foreground subjects including people, animals, products, and objects against complex backgrounds. Progress is reported in real time as the model initializes and processes your image. After processing, a comparison slider lets you swipe between the original and the result to evaluate quality before downloading.
Your images are never sent to any server. The entire AI inference pipeline — model weights, pixel processing, output generation — runs in your browser tab using WebAssembly. This is meaningfully different from cloud-based background removal services that require uploading photos to a remote server.
Why does the first run take a while to start?
The AI model (~50 MB) is downloaded from a CDN on first use. After that, it's cached and starts instantly.
Does it work on images with complex or messy backgrounds?
The model performs well on most photos. Complex backgrounds with many similar colors to the subject (e.g., a person wearing the same color as the background) may produce less accurate results.
What format is the output?
The output is always a transparent PNG file, regardless of the input format.