image-segmentation · vision · WebGPU (fp32) · ~170 MB
Give it a photo and it decides, pixel by pixel, what's foreground and what's background — returning a soft alpha matte you can drop onto transparency, a colour, or a new scene. It's the "remove background" button, running entirely on your device. Nothing is uploaded.
| Model | briaai/RMBG-1.4 |
|---|---|
| Task / pipeline | image-segmentation — a single foreground/background matte |
| Architecture | IS-Net dichotomous segmentation → one alpha value per pixel |
| Precision | fp32 on WebGPU; q8 (8-bit) on the WASM fallback |
| Backend | WebGPU (fp32); falls back to WebAssembly where WebGPU is absent |
| Download | ~170 MB (fp32) / smaller on the q8 fallback, cached after first load |
| License | Non-commercial (BRIA RMBG-1.4); commercial use needs a BRIA agreement |
| Web APIs | WebGPU, Web Workers, Canvas 2D, FileReader — WebGPU is the only non-Baseline piece, and the page says so if it's missing |
Pick or drop a photo, then remove the background. The result sits on a checkerboard so you can see exactly where it made the image transparent.
The model never sees "a cat" — it outputs an alpha matte: a number from 0 (background) to 255 (subject) for every pixel. The cutout above is just the photo with that matte stamped into its transparency. The soft, in-between values along the boundary are where the model expresses uncertainty — wisps of fur, hair, motion blur — and they're what make a cutout look real instead of cut with scissors.
| matte resolution | – |
|---|---|
| foreground coverage | – |
| soft-edge pixels | – |
"Remove the background" used to mean a round-trip to a paid API or twenty minutes in Photoshop. RMBG does it in a web page, on-device, in a second or two — which changes what you can build:
RMBG-1.4 uses a custom architecture, so the model card's documented path loads the model and processor directly and reads the matte tensor:
import { AutoModel, AutoProcessor, RawImage } from "@huggingface/transformers";
const model = await AutoModel.from_pretrained("briaai/RMBG-1.4", {
device: "webgpu", dtype: "fp32", config: { model_type: "custom" },
});
const processor = await AutoProcessor.from_pretrained("briaai/RMBG-1.4");
const image = await RawImage.read(imageURL);
const { pixel_values } = await processor(image);
const { output } = await model({ input: pixel_values });
// output[0] is a 0–1 foreground map. Scale to 0–255 and resize to the original photo:
const matte = (await RawImage.fromTensor(output[0].mul(255).to("uint8")))
.resize(image.width, image.height);
// then copy `matte` into the image's alpha channel → the cutoutThis page runs that in a Web Worker and copies the matte into the photo's alpha on a canvas, so the interface stays responsive and the composite (transparency, colour, or a new scene) is all client-side.