Remove Image Background — Free, In-Browser, No Signup

The tool runs the ISNet (Highly Accurate Dichotomous Image Segmentation) neural network — the same architecture published by Qin et al. at ECCV 2022 — through ONNX Runtime Web inside your browser. The model is loaded and orchestrated by @imgly/background-removal, an open-source wrapper from IMG.LY. The browser pulls the model weights once from the IMG.LY CDN, caches them in IndexedDB, and from then on every image is processed without any network call on the photo bytes themselves.

A run on one image goes through five deterministic steps:

1. Validate. The file must be JPG, PNG, or WebP, under 50.0 MB, and at most 4,096 × 4,096 px. Rejected files leave a specific reason in the queue rather than failing silently.
2. Decode. Your browser's native image decoder reads the bytes into an ImageData buffer. No third-party decoder is loaded.
3. Pre-process. ISNet was trained on 1024×1024 px input. The runtime resizes the longest side to 1024 px, normalises channels to a zero-mean distribution, and reshapes into the BCHW tensor layout expected by the model.
4. Inference. ONNX Runtime evaluates the network on either the WebGL or WASM execution provider depending on your browser. The output is a single-channel alpha mask the same size as the resized input, with values 0 (background) to 1 (subject).
5. Composite & encode. The alpha mask is upsampled with bilinear interpolation back to the original image dimensions, multiplied with the RGB channels of the source, and the result is encoded as a PNG (transparent), WebP (transparent), or JPEG (flattened on the colour you pick). The encoded Blob is wrapped in an object URL and offered as a download.

The first-run download time depends on which model variant you pick and your network bandwidth. The page exposes two equivalent estimators so you can sanity-check the wait before clicking. The closed-form throughput estimator computes seconds = bytes ÷ (Mbps × 1,048,576 ÷ 8) + 1.5 s for TCP/TLS overhead. The lookup estimator interpolates a piecewise table calibrated against real SLT-Mobitel and Dialog measurements at 5, 25, 50, 100, 250, and 500 Mbps. Both agree within ±5% at every table speed — for the default model (44 MB) on a 50 Mbps line, throughput predicts 8.9 s and lookup predicts 9.0 s.

Processing time per image scales linearly with megapixels above the 1024 px training size. On a 2022 MacBook Air (M2) running the CPU execution provider, a 1 MP image finishes in about 4 seconds. On a mid-range Android phone, the same image takes 9–12 seconds. WebGL acceleration cuts that to ~1 second on GPUs that support 16-bit floats, but @imgly defaults to CPU for maximum compatibility.

Each variant of the ISNet model is a different trade-off between download size, inference speed, and edge quality. The default — isnet_fp16 — covers ~95% of users. Switch only if you have a specific reason.

• IMG.LY — @imgly/background-removal (runtime, MIT/AGPL dual licence)
• IMG.LY — Background Removal showcase and model card
• Qin et al., Highly Accurate Dichotomous Image Segmentation (ECCV 2022) — the ISNet paper
• ONNX Runtime Web — execution providers (WASM, WebGL)
• MDN — HTMLCanvasElement.toBlob (browser-native PNG/WebP/JPEG encoder)

The runtime version, model variants, and cited paper were last cross-checked on 2026-05-11. The page is reviewed whenever @imgly/background-removal ships a new model card.