Background removal here means running the ormbg ONNX model (Apache-2.0) inside your tab. It is not an upload API. The interesting choice is how the browser executes that model: WebGPU when possible, otherwise CPU with a WASM runtime path.
Prefer WebGPU when it is available
- Open Capability and check whether this browser reports WebGPU.
- If a GPU adapter exists, LocalPix prefers WebGPU for inference speed on typical laptop/desktop GPUs.
- If the adapter is missing (common on some locked-down corporate images, older browsers, or remote desktop setups), forcing WebGPU cannot help — the tool should fall through rather than hang.
When WASM / CPU is the right path
- No GPU adapter — WebGPU is skipped; CPU/WASM keeps processing local.
- WebGPU init fails after pick — fallback exists so a single broken GPU path does not strand the file on a server (there is no server cutout).
- You are verifying privacy — WASM still runs in-process. Model weights may download from allowed hosts after you choose a file; your photo should not POST to the LocalPix origin. See Verify local.
WASM is not “more private” than WebGPU by itself. Both keep pixels in the tab. WASM is the compatibility lane.
What downloads (and what must not)
- Before file pick: page CSS/JS only. Large ormbg weights should not start eagerly.
- After file pick: you may see ONNX / WASM / CDN requests (e.g. Cloudflare R2 or Hugging Face fallback for ormbg). That is model delivery, not your photo leaving the device.
- Download of the cutout: a local
blob:URL. No multipart upload of the result to LocalPix.
Practical checklist
- Open Remove background.
- Optional: confirm backend messaging (WebGPU vs WASM/CPU) once inference starts.
- If cutout is slow on WASM, try a smaller source image or a machine with WebGPU — do not switch to an upload-based service if privacy is the reason you chose LocalPix.
- HEIC inputs: convert with HEIC to JPG (libheif-js WASM on demand) first when the browser cannot decode HEIC natively.
FAQ
Does LocalPix upload my photo for background removal?
No. Inference runs in your browser with ormbg. Your image bytes are not uploaded to LocalPix servers. Model weights may download after you choose a file.
When does LocalPix use WASM instead of WebGPU?
WebGPU is preferred when a GPU adapter is available. If there is no adapter, or WebGPU fails, the tool falls back to CPU/WASM so cutout can still run locally.
When is the ormbg model downloaded?
Only after you pick an image. Opening the page alone should not start the large ONNX download. Details: Licenses.
Related
Remove background · Capability · Verify local · Compress PNG for email