Show HN: Remove-bg – open-source remove background using WebGPU

Original Article ↗ Hacker News Discussion ↗

A new open-source tool uses WebGPU to run a neural-network model entirely in the browser to remove image backgrounds, avoiding server-side processing and enabling fully local workflows. Commenters praise how close it comes to commercial products with relatively little code, but report frequent crashes, high memory use, large model downloads, and limited browser support, highlighting how immature WebGPU and browser-based AI still are. The thread also surfaces concerns about restrictive model licenses and points to alternative background-removal models and libraries with more permissive terms or better quality.

Overall impressions

  • Many commenters find the in-browser background remover impressive, especially given how little custom code is needed with WebGPU + Transformers.
  • Several note that commercial tools (Photoshop, macOS Preview, remove.bg, etc.) still give better and more controllable results, but praise how close this open-source, browser-only tool gets.

Image quality & use cases

  • Works very well for many photos, including tricky subjects like animals, people with hair, and complex tack/clothing.
  • Fails badly on some non-photo images (e.g., charts/plots), sometimes removing the key content while keeping background.
  • Some users see artifacts or warping on the subject after background removal.
  • Quality is widely viewed as driven mainly by the model; multiple people hope for higher-quality or alternate models.

Browser, GPU & stability issues

  • WebGPU support is a major friction point:
    • Chrome/Chromium on Linux often requires flags like --enable-unsafe-webgpu --enable-features=Vulkan.
    • Firefox largely doesn’t work (WebGPU not enabled by default; some users get tab crashes or transparent output).
    • Some browsers (Arc, certain Chromium setups) freeze, crash tabs, or trigger OOM killer; reports of several GB RAM usage.
  • Developer added better error banners and troubleshooting instructions but detection/fallback logic is still fragile.

Model, licensing & dependencies

  • Uses the BRIA RMBG-1.4 model; some question whether the demo complies with the evaluation-only, non-distribution license.
  • Debate over whether streaming model weights to the browser counts as “distribution.”
  • Others note that many newer or stronger models have restrictive licenses; suggest open-source options (e.g., U2-Net, BiRefNet, InSPyReNet, isnet).
  • Complaints about heavy npm dependency tree and use of an older React canary, attributed to quick prototyping.

Data usage, offline behavior & privacy

  • Model download is large (~176 MB total page transfer for some users), surprising those on metered connections.
  • Requests to:
    • Show or confirm model size before loading.
    • Warn about memory usage.
    • Possibly allow an explicit “offline” mode guarantee.

Naming & ecosystem context

  • Name collides with an existing commercial service; some see confusion risk, though the author treats this as an experimental project.
  • Thread surfaces many related tools: CLI background removers, other WebGPU/ONNX-based libraries, and API-based services, highlighting a growing ecosystem around on-device background removal.