4xNomos2 OTF ESRGAN

4× photo upscaler that also cleans up JPEG compression and noise.

An ESRGAN upscaler fine-tuned from Real-ESRGAN x4plus on the Nomos-v2 photo dataset with on-the-fly degradations, so it removes compression and noise while it upscales.

Download ONNX View repository Paper

Super-resolution Computer Vision

Scale: 4×
Parameters: 16.7M
CPU time per 256 px tile: 8.4s
ONNX file: 67MB

Example output

Drag the handle to compare. Produced by running this model.

After: Input: a 120 px street photo saved as a low-quality JPEG, enlarged with bicubic resizing. Output: the model's 4× result. — Input: a 120 px street photo saved as a low-quality JPEG, enlarged with bicubic resizing. Output: the model's 4× result.

Before: Input: a 120 px street photo saved as a low-quality JPEG, enlarged with bicubic resizing. Output: the model's 4× result. — Input: a 120 px street photo saved as a low-quality JPEG, enlarged with bicubic resizing. Output: the model's 4× result.

Technical details

Architecture: ESRGAN (RRDBNet, 64 features, 23 blocks)
Fine-tuned from: Real-ESRGAN x4plus
Training data: Nomos-v2, 6,000 photos, on-the-fly degradations
Training: 246,000 iterations
ONNX build: FP32, opset 17, any image size
Input: RGB, NCHW, values 0–1
Benchmark: ONNX Runtime CPU, Apple M2 Pro
Original release: 22 June 2024

Good for

Compressed photos
Old phone pictures
Images saved from the web

Run it in Flutter

This model runs as-is with our open-source flutter_super_resolution package (MIT), which tiles large images and uses Core ML on Apple devices when available. It works on iOS, Android (ARM devices), macOS and Windows.

pubspec.yaml · Dartgeneral4xhigh.onnx

dependencies:
  flutter_super_resolution: ^1.1.1

// Load the model, then upscale a ui.Image
final upscaler = FlutterUpscaler(tileSize: 128, overlap: 8);
await upscaler.initializeModel('assets/models/general4xhigh.onnx');

final result = await upscaler.upscaleImage(
  image,
  scale: 4,
  onProgress: (progress, message) => print('$message ${(progress * 100).round()}%'),
);

upscaler.dispose();

Package on GitHub All models that work with it