image_guard 0.1.0-dev.2

On-device NSFW (18+) image detection and compression to an exact byte limit and resolution, powered by a Rust core.

image_guard

On-device NSFW (18+) image detection and compression to an exact byte limit and resolution for Flutter, powered by a Rust core (via flutter_rust_bridge).

The AI check runs on the device. Images are never uploaded anywhere to be checked.

Features

NSFW detection

• On-device MobileNetV2 classifier (GantMan/nsfw_model, MIT). It runs with tract, a pure-Rust inference engine, so no TensorFlow/ONNX Runtime binaries are added to your app.
• Five classes: drawings, hentai, neutral, porn, sexy.
• Verdict: safe / uncertain / unsafe, with thresholds you can configure.
• The check runs on the full-quality image before compression. Unsafe images are not compressed.

Compression

• Exact size limit: output is always <= maxBytes, or a clear cannotMeetTarget error. It never quietly returns a file that is too big.
• Resolution limit: maxWidth / maxHeight. The aspect ratio is kept and images are never upscaled.
• Never makes small images worse: a 100 KB photo with a 500 KB limit comes back unchanged.
• Picks the highest quality that fits. Resolution is reduced only if needed.
• Fixes EXIF rotation and strips EXIF/GPS metadata.
• Reads JPEG, PNG, WebP, GIF and BMP, and writes JPEG or PNG.
• Rejects broken, oversized (> 100 MP) and malicious input without crashing.

Usage

import 'package:image_guard/image_guard.dart';

void main() async {
  WidgetsFlutterBinding.ensureInitialized();
  await SafeImage.initialize(); // optional: loads the model up front (~0.2 s)
  runApp(const MyApp());
}

final result = await SafeImage.process(
  SafeImageSource.file(xfile.path),
  options: SafeImageOptions(
    maxBytes: 50.kb,   // user-chosen size
    maxWidth: 1080,    // user-chosen resolution (optional)
    maxHeight: 1080,
  ),
);

switch (result.verdict) {
  case Verdict.unsafe:
    showError('18+ content is not allowed'); // result.image is null
  case Verdict.uncertain:
    // Your policy: reject, or send to manual review.
    // (A case must not be empty: in Dart an empty case falls through to the next one.)
    showError('This photo needs review');
  case Verdict.safe:
    await api.upload(result.image!.bytes);   // guaranteed <= 50 KB
}

print(result.safety); // SafetyReport(safe, nsfwScore: 0.012, porn: 0.004, ...)

Other calls:

final report = await SafeImage.classify(source);   // NSFW check only
final small  = await SafeImage.compress(source);   // compression only (no NSFW check!)
await SafeImage.dispose();                         // free model memory

Safety options

Option	Default	Meaning
rejectThreshold	0.7	nsfwScore >= this → unsafe
reviewThreshold	0.3	nsfwScore >= this → uncertain
suggestiveWeight	0.5	How much sexy (swimwear, lingerie) counts. 0 = ignore, 1 = same as explicit
compressUnsafe	false	Also compress unsafe images

nsfwScore = porn + hentai + sexy × suggestiveWeight. SafetyOptions.strict is a stricter preset.

Accuracy: about 92 % validation accuracy on the upstream dataset. Like every classifier it makes mistakes (false positives and false negatives). For user-generated content, combine it with server-side moderation and/or reporting. This is not a CSAM detection tool.

Errors

try {
  await SafeImage.process(source, options: options);
} on SafeImageException catch (e) {
  switch (e.code) {
    case SafeImageErrorCode.invalidOptions:   // bad maxBytes / quality / thresholds
    case SafeImageErrorCode.unsupportedFormat: // e.g. HEIC, PDF
    case SafeImageErrorCode.corruptImage:
    case SafeImageErrorCode.cannotMeetTarget:  // limit too small for minDimension/minQuality
    case SafeImageErrorCode.modelLoadFailed:
    default:
  }
}

Compression options

Option	Default	Allowed
maxBytes	50 * 1024	1 KB – 100 MB
maxWidth / maxHeight	null (no limit)	1 – 16384, and >= minDimension
format	jpeg	jpeg, png
minQuality / maxQuality	40 / 90	1 – 100, min <= max
minDimension	64	1 – 4096
keepOriginalIfFits	true
stripMetadata	true

HEIC on iPhone

HEIC is not decoded yet. When you use image_picker, pass imageQuality (or maxWidth). iOS then converts the photo to JPEG before returning it.

Model & license

The bundled model assets/models/nsfw_mobilenet_v2_140_224.nnef.tar is converted from GantMan/nsfw_model release 1.2.0 (MIT License, Copyright (c) 2020 The nsfw_model Developers). See NOTICE and tool/convert_model.py.

Platform support & requirements

Platform	Minimum
Android	arm64-v8a, armeabi-v7a, x86_64, x86
iOS	13.0 (15.0 with Xcode 27+)
macOS	10.15 (for development)

Rust toolchain required (for now). The native library is compiled while your app builds:

1. Install Rust: https://rustup.rs
2. Android: the Android NDK (Android Studio → SDK Manager → NDK).
3. iOS: Xcode.

The first build compiles Rust and takes a few minutes. Later builds are cached. Precompiled binaries are planned, and then Rust will no longer be needed.

Xcode 27+: Xcode 27 no longer builds for iOS 13. Set iOS 15 in ios/Podfile:

platform :ios, '15.0'

post_install do |installer|
  installer.pods_project.targets.each do |target|
    flutter_additional_ios_build_settings(target)
    target.build_configurations.each do |config|
      config.build_settings['IPHONEOS_DEPLOYMENT_TARGET'] = '15.0'
    end
  end
end

Also set iOS Deployment Target = 15.0 for the Runner target in Xcode.

Development

cargo test --manifest-path rust/Cargo.toml          # Rust unit + scenario + property tests
cargo build --release --manifest-path rust/Cargo.toml
flutter test                                         # Dart tests (uses the host build above)
cd example && flutter test integration_test -d <device>
flutter_rust_bridge_codegen generate                 # after changing rust/src/api