Cutout model · Fast (U²-Net small)
Lightweight general subject cutout for quick transparent PNG exports.
Size: 4.6 MB · License: Apache-2.0Used by: AI Cutout · ID Background · Spotlight
Model setup
Local model manager
Prepare segmentation, face landmark, object detection, OCR, and runtime resources in this browser before running local AI tools.
ONNX model packs (cutout / upscale / erase)
These packs power AI cutout, AI upscaling, and smart erase. Prepare them here, or let each tool prepare its model on first use. Models download once, stay cached in this browser, and images never leave your device.
Site owner tip: run npm run models:fetch in the repo to bundle the models with the site so visitors never need an external download.
Available
Cutout model · Quality (Silueta)
Balanced general subject cutout with clearly better edges for people, products, and pets.
Size: 42.6 MB · License: Apache-2.0Used by: AI Cutout · ID Background
Cutout model · Max (ISNet general)
Highest quality general cutout including hair and fine edges. Large download.
Size: 170.0 MB · License: Apache-2.0Used by: AI Cutout
Upscale model (Real-ESRGAN x4v3)
AI 2x and 4x image upscaling with detail reconstruction for photos and graphics.
Size: 4.8 MB · License: BSD-3-ClauseUsed by: AI Upscale
Erase model (MI-GAN)
Generative fill for removing watermarks, passers-by, and objects with natural reconstruction.
Size: 27.0 MB · License: MITUsed by: Remove Objects
Browser compatibility
Checking browser capabilities…
Local AI resources
Preload stores model runtimes and weights in this browser cache. Tools read cached model resources first during processing.
MediaPipe vision runtime
Runs browser-local segmentation models and face landmark models.
Resource size: 21.7 MBRuntime path: CPU / WebAssemblyTools: Background blur, Subject spotlight, Remove background, ID photo background, Portrait toolsPreload
Portrait segmentation model
Separates people from the background for background blur, subject spotlight, transparent PNG, and ID photo backgrounds.
Resource size: 0.24 MBRuntime path: CPU / WebAssemblyTools: Background blur, Subject spotlight, Remove background, ID photo backgroundPreload
Face landmark model
Reads face geometry for makeup placement, portrait tone, face privacy blur, teeth, eyes, and hair colour preview.
Resource size: 3.59 MBRuntime path: CPU / WebAssemblyTools: Makeup preview, Hair colour preview, Portrait enhance, Face privacy blur, Teeth whiteningPreload
Hair segmentation model
Detects the hair region directly for hair colour preview and stronger face protection.
Resource size: 0.75 MBRuntime path: CPU / WebAssemblyTools: Hair colour previewPreload
Detailed portrait part segmentation model
Separates hair, face skin, body skin, clothing, and accessories for fine hair colour and portrait effects. Downloads on demand at first use and stays cached.
Resource size: 15.6 MBRuntime path: CPU / WebAssemblyTools: Hair colour preview, Portrait effectsPreload
Object detection model
Runs browser-local object detection for counting, inspection, labels, shelves, parking spaces, and review-style vision tools.
Resource size: 4.50 MBRuntime path: CPU / WebAssemblyTools: Shelf audit, Warehouse labels, Parking space detection, Paper countingPreload
Universal YOLO nano model
Runs browser-local YOLO object detection for common countable objects such as cups, bowls, bottles, tableware, parcels, people, and vehicles.
Resource size: 7.50 MBRuntime path: CPU / WebAssemblyTools: Universal object counting, Local vision countingPreload
Optional tableware ONNX model pack
Detects countable table objects such as chopsticks, bamboo sticks, food picks, plates, saucers, bowls, and cups with a fixed vocabulary.
Resource size: 28.0 MBRuntime path: CPU / WebAssemblyTools: Universal object counting, Stick counting, Tableware countingPreload
Optional stick ONNX model pack
Prioritises visible end-face candidates for bamboo sticks, food picks, sample sticks, and thin marker rods.
Resource size: 7.50 MBRuntime path: CPU / WebAssemblyTools: Stick tray countingPreload
RapidOCR PP-OCRv4 models
Runs PaddleOCR-based text detection and recognition for higher accuracy, especially on Chinese text.
Resource size: 15.0 MBRuntime path: CPU / WebAssemblyTools: OCR Image, Document workflowsPreload
Browser OCR runtime
Loads the same-origin Tesseract.js runtime, the LSTM core, and fast English and Simplified Chinese language data.
Resource size: 24.4 MBRuntime path: CPU / WebAssemblyTools: OCR Image, Document workflowsPreload