How It Works

AI Background Removal

ISNet neural network via ONNX Runtime WebAssembly — enterprise-grade AI running privately in your browser.

ISNet Neural Network

Intermediate Supervision Network — a state-of-the-art deep learning architecture for salient object detection, trained on 5,470 high-quality images.

  • Multi-scale feature extraction for precise edge detection
  • Deep supervision through intermediate layers
  • Encoder-decoder architecture for fine detail
  • Trained on DIS5K dataset with pixel-perfect masks

ONNX Runtime WASM

Open Neural Network Exchange Runtime compiled to WebAssembly — near-native AI inference with no server required.

  • SIMD instructions for parallel tensor operations
  • Optimized convolution and pooling operators
  • Memory streaming for large images
  • Identical results across all modern browsers

AI processing pipeline

1

Image Input

Decode & normalize pixels

2

Feature Extract

CNN layers detect patterns

3

Segmentation

Generate alpha mask

4

Mask Apply

Separate foreground

5

Encode Output

PNG/WebP with alpha

Model variants

ModelPrecisionSizeSpeedBest for
ISNet FullFP32 (32-bit)~20 MB15–30 sProfessional work, complex subjects
ISNet FP16FP16 (16-bit)~10 MB8–15 sBalanced quality & performance
ISNet Quint8INT8 (8-bit)~5 MB3–8 sBulk processing, quick results

Technical specifications

Input formatsJPEG, PNG, WebP, BMP, TIFF
Output formatsPNG (alpha), WebP (alpha), JPEG
Max image size4096×4096 px
Model load (first visit)10–30 s
Model load (cached)< 1 s
Processing 1024×7683–20 s depending on model
Memory usage200–500 MB
AI library@imgly/background-removal

ISNet runs locally

The model is downloaded once and cached in your browser — subsequent runs are instant with no network dependency.

3 quality levels

Choose Quint8 for speed, FP16 for balance, or Full precision for professional-grade results with complex subjects.

Transparent PNG output

Output preserves the alpha channel so your cutout works on any background without fringing or artifacts.