Background Remover

Background Remover runs the RMBG-1.4 segmentation model directly in your browser. The model is loaded through @huggingface/transformers (the JavaScript port of Hugging Face's Python library) and executed by onnxruntime-web on a Web Worker so the UI stays responsive while inference runs.

On the first use the worker downloads roughly 44 MB of quantised ONNX weights from the Hugging Face CDN and caches them in IndexedDB. Every subsequent run is offline — the weights are pulled from disk, the WebAssembly runtime is reused, and a 1024-pixel image typically segments in two to four seconds on a modern laptop CPU.

Before inference the original file is normalised: EXIF orientation is read straight from the JPEG bytes (so a phone photo flagged as orientation 6 is rotated upright before pixels are sent into the model) and the image is drawn into a Canvas to give the worker a flat RGBA buffer. The buffer is transferred — not copied — into the worker to keep memory flat for large images.

The model returns a single-channel alpha mask the same size as the input. Toolnest then composites the mask back onto the original pixels, producing a true transparent PNG. You can keep the cutout transparent, fill the background with a solid colour or one of the built-in CSS gradients, drop in your own background image, or apply a tunable Gaussian-style blur to the original backdrop.

Edge refinement is optional. The Edge Smoothing slider applies a small dilate/erode pass to the alpha channel to reduce halos around hair and fur, and Detail Enhancement boosts local contrast where the alpha is partial. Neither is a magic fix — RMBG-1.4 is trained on natural photos, so heavy motion blur, glass, transparent fabric, or backgrounds that exactly match the subject's colour will still need touch-up in a pixel editor.

Output is always 8-bit PNG (with optional WebP) and you control the export size: original, half, double, or a custom width and height. There is no upload step and no API key — the image, the model weights, and the alpha buffer all live inside the tab. Closing the tab frees everything; the model stays cached for next time.

For best results give the tool a sharp, well-lit subject on a contrasting background. The 1024x1024 input resolution means the model down-samples very large images during inference; for 4K product photos it is normal to lose a pixel or two of edge precision, which the smoothing slider hides reasonably well.