PICORN Remove Background
Background remover in the Browser/Client side.

Try the live demo · Run locally · Smart refine

PICORN removes image backgrounds without sending the image to a server. The model, inference runtime, alpha correction and PNG export all run in the browser.

The V12 ONNX model is 1.2 MB (1,209,207 bytes) and has 524,198 inference parameters. It keeps the V10 hybrid edge stage and the white-graphic refinement, and adds a zoomed second pass, confident-alpha snapping and foreground colour estimation. Inference uses WebGPU when supported, with a WASM fallback. See the V12 notes.

V12 model update

V12 swaps the from-scratch encoder for an ImageNet-pretrained MobileNetV4-Small-050 and keeps the V11 decoder. Weights are stored as FP16 and cast to FP32 at load, so the file is smaller than V11 while all arithmetic stays FP32. On the original 6,489-image holdout at 512 px:

Metric	V11	V12
Intersection over Union	0.617	0.773
Mean absolute alpha error	0.073	0.036
Boundary F1	0.395	0.596
Model file	1,417,525 B	1,209,207 B
CPU, 512 px, 4 threads	77.0 ms	76.7 ms

What you can do

• Drop in a PNG, JPG or WebP. Nothing is uploaded.
• Choose 320, 384 or 512 px inference profiles for different devices.
• Inspect the edge on a checkerboard, solid color or gradient.
• Paint green to restore a missing region or red to remove a leftover region.
• Download the cutout, a composite or the alpha matte as PNG.
• Run the same static build on GitHub Pages, a CDN or your own machine.

Why it is small

PICORN does not put a large general-purpose vision backbone in the browser. Its ONNX graph combines a compact semantic encoder with dynamic foreground and background prototypes, multi-scale texture signals and a shallow full-resolution detail path. It predicts a soft alpha matte, not just a binary object mask.

The result is one model file with 990 ONNX nodes and 146 initializers. There is no Python service behind the demo.

V10 hybrid edge update

V10 separates two jobs that do not need the same tool:

1. The compact neural model decides what the foreground object is and produces a soft alpha matte.
2. Eight four-neighbour WebGPU passes propagate only uncertain alpha values between pixels with similar RGB values.
3. A quarter-resolution fast guided filter places the soft transition back on the full-resolution image edge.
4. Confident alpha anchors, the foreground/background class boundary and thin structures without a nearby safe core are protected.

The refiner has no learned weights and does not change the 1,417,388-byte ONNX artifact. Chromium uses WGSL compute shaders; other browsers run the same deterministic CPU reference implementation.

On the complete 6,489-image validation set:

Metric	V9 model	V10 hybrid	Change
Intersection over Union	0.59294	0.59412	+0.00118
Mean absolute error	0.07449	0.07423	-0.00026
Mean squared error	0.05174	0.05135	-0.00040
Boundary F1	0.38432	0.39194	+0.00762

The improvement is positive on DIS5K, DUTS-TE and P3M-10K individually. A real Chrome WebGPU run matched the CPU reference within 1.2e-7; at 512 px the warm edge stage averaged 11.25 ms on the validation workstation. See the V10 research and benchmark notes for the paper trail, ablation results and limitations.

V9 model update

V9 improves the matte without adding layers, parameters or browser code. It is the same compact graph, fine-tuned once more at a low learning rate.

On the full 6,489-image validation set (DIS5K, DUTS-TE and P3M):

Metric	V8	V9
Intersection over Union	0.59233	0.59294
Mean absolute error	0.07515	0.07449
Boundary F1	0.38206	0.38432

The model update is deliberately small. V9 improves the aggregate score, edges and alpha error while keeping the 1,417,388-byte artifact unchanged. See the V9 research notes for the per-dataset results and rejected experiments.

How the refine brush works

The brush is more than a round paint stamp:

1. The current alpha matte supplies confident foreground and background anchors.
2. A green or red stroke becomes a local trimap hint.
3. Color, gradient and Shannon-entropy differences decide how far that hint may travel through the connected region.
4. Small mask islands and enclosed holes can be corrected as complete components.
5. A guided filter pulls the new soft alpha back onto the visible image edge.

Only the narrow painted core is a hard instruction. The grown selection stays a soft proposal, and the correction is limited to a local band. This prevents one small hint from replacing most of an otherwise good mask.

Run it locally

You need Node.js 22 or newer.

git clone https://github.com/cueqzapper/PICORN-Background-Remover.git
cd PICORN-Background-Remover
npm install
npm run dev

Open http://127.0.0.1:4185/.

For a production build:

npm test
npm run check
npm run build
npm run preview

Browser support

Browser	Preferred path	Fallback
Chrome / Edge	WebGPU	WASM
Firefox	WASM	—
Safari	WebGPU where available	WASM

The page becomes interactive before the runtime is warm. On first use, the browser downloads and caches the model and the relevant ONNX Runtime files. Later runs reuse the browser cache. Modern Chromium browsers use the smaller JSPI WebGPU runtime; older WebGPU implementations retain a compatible fallback.

Project layout

public/models/                 PICORN ONNX model
public/demo/                   CC0 demo images
src/browser-runtime.js         ONNX loading, preprocessing and hybrid routing
src/hybrid-refine-webgpu.js    WebGPU edge graph and guided alpha passes
src/hybrid-refine.js           Deterministic CPU reference and fallback
src/interactive-matting.js     Trimap, geodesic growth and guided refinement
src/main.js                    Demo interaction, rendering and PNG export
tests/                         Runtime and smart-refine regression tests

Limits

The model deliberately trades broad model capacity for size and client-side speed. Extremely crowded scenes, transparent glass, motion blur and subjects whose color matches the background can still need a refine stroke. The Quality profile improves thin structures, but it also does more work.

Licenses

• Browser source code: MIT
• Included PICORN ONNX weights: separate model terms
• Included demo images: CC0 1.0
• ONNX Runtime Web is installed from npm and retains its MIT notice.

If you want to use the weights in a commercial product or hosted service, contact PICORN for a commercial model license.