AI Background Remover — Transparent PNG at Full Resolution, No Upload
Remove the background from any photo entirely in your browser. The tool runs BiRefNet_lite — a dichotomous image segmentation model — as an ONNX fp16 graph inside a Web Worker via ONNX Runtime Web, accelerated by WebGPU with fp16 shader support (required — Chrome / Edge 113+, Safari 26+ on a device with a modern GPU). The page detects WebGPU fp16 support up front; unsupported setups receive an instant notice and nothing is downloaded. Inference runs at the model's native 1024×1024 input with ImageNet normalisation; the resulting mask is bilinearly upscaled and applied as an alpha channel to the original pixels, so the output PNG is your photo at its full source resolution. On first use, ~140 MB total downloads (the ~115 MB model plus the ~26 MB inference runtime), split into ≤20 MB chunks to fit Cloudflare Pages limits, reassembled in the browser, and persisted to Cache Storage — subsequent uses skip the download entirely.
How to remove a background
- Drop your photo onto the tool or click to browse.
- The page checks for WebGPU fp16 support up front — if your browser or device does not support it, you see an instant notice and nothing is downloaded.
- On first use with a supported browser, ~140 MB total (the ~115 MB model plus the ~26 MB inference runtime) downloads in chunks and persists to your browser's Cache Storage — this happens once.
- The model infers a foreground mask at 1024×1024, which is upscaled and applied as an alpha channel to your original pixels.
- Download the result as a transparent PNG at your photo's full resolution.
Common use cases
- Isolating products for e-commerce listings against a transparent or solid background.
- Removing backgrounds from headshots for professional profiles or presentation slides.
- Preparing cut-out subjects for compositing in design tools without round-tripping through a cloud service.
- Creating sticker-style transparent PNGs from photos for messaging apps or digital collages.
Frequently asked questions
Is my photo uploaded to a server?
No. The AI model runs inside a Web Worker in your browser — all computation is local. DevTools Network shows no upload requests when you process an image: the model was downloaded to your device on first use and is read from Cache Storage on every subsequent run.
Why does it download ~140 MB the first time?
The download totals ~140 MB: the BiRefNet_lite model is an ONNX fp16 graph of 114,538,221 bytes (~115 MB), plus the ~26 MB ONNX Runtime Web inference engine. Both are split into ≤20 MB chunks (Cloudflare Pages file limit), fetched same-origin, reassembled in the browser, and written to Cache Storage via the CacheStorage API. Future sessions read from the local cache — no re-download.
Is full resolution really free, with no watermark?
Yes. The output PNG carries the alpha channel applied to your original pixels at their full source dimensions. There is no server-side processing tier, no paid plan, and no watermark. The site is funded by unobtrusive ads on the page.
How does it handle hair and fine edges?
BiRefNet_lite uses a dichotomous image segmentation architecture specifically designed for fine-structure subjects. Results on hair, fur, and thin fabric are generally better than coarser segmentation approaches, but quality varies by image — lighting, contrast, and background complexity all affect the mask. No quality guarantees are made for any specific photo.
Which browsers are supported?
WebGPU with fp16 shader support is required. Chrome 113+, Edge 113+, and Safari 26+ on a computer or phone with a modern GPU meet this requirement. The page detects support on load — unsupported setups (Firefox, software renderers, older GPUs, headless or virtualised environments) receive an upfront notice and nothing is downloaded.
How the background removal works
BiRefNet (Bilateral Reference Network) is a dichotomous image segmentation architecture designed for fine-structure subjects — hair strands, thin fabrics, detailed outlines — where coarser segmentation approaches produce blocky or ragged edges. The model was released by ZhengPeng7 on GitHub under the MIT licence; the fp16 ONNX export used here (114,538,221 bytes) is published as onnx-community/BiRefNet_lite-ONNX on Hugging Face under the same licence. ONNX Runtime Web (microsoft/onnxruntime, MIT) handles the browser inference via the WebGPU backend. The fp16 model requires the optional WebGPU shader-f16 feature, so a modern GPU and a WebGPU-capable browser (Chrome/Edge 113+, Safari 26+) are required. The output is a W3C PNG with an alpha channel — the only lossless raster format with full per-pixel transparency.
- Runs 100% in your browser — photo never leaves the device
- Full-resolution transparent PNG output at original source dimensions
- BiRefNet_lite ONNX fp16 model — designed for fine edges and hair
- WebGPU with fp16 shader support required (Chrome/Edge 113+, Safari 26+ on a modern GPU)
- ~140 MB on first use (the ~115 MB model plus the ~26 MB inference runtime), persisted to Cache Storage
- No account, no watermark, no file-size cap
Free. No signup. No file uploads. Ads via AdSense (consent required).