Light Converter
Batch background removal that runs entirely in the browser — no upload, no account.
- Period
- 2026 — now
- Platforms
- WebAI
- Stack
- AstroSvelteTypeScriptonnxruntime-web
- Links
- Source ↗
A neural segmentation model running on the visitor's own GPU, so the images never leave the machine they were opened on.
The problem
Every background remover on the web asks you to upload your images to someone else’s server. For anything sensitive — documents, product shots under embargo, photographs of people — that is the whole objection, and no privacy policy answers it.
Constraints
If the images must never leave the device, the model has to run on the device. That rules out the usual architecture and puts a real neural network in the browser, on hardware that is unknown until the page loads.
What we built
BiRefNet runs client-side through onnxruntime-web, on WebGPU where it is available and falling back to WASM where it isn’t — so the tool degrades in speed rather than failing outright. A guided-filter pass refines the matte at the edges, which is where cheap segmentation usually gives itself away on hair and thin detail.
Batch processing runs in the same session without a round trip, because there is no round trip.
What shipped
Open source, and honest about its claim: there is no upload endpoint to audit, because there is no upload. It is also where the Astro and WebGPU work that shows up elsewhere in the practice was first tried.