An app for removing compression and noise from photos

This is not another "AI upscaler" but a compression artifact remover and denoiser using public models such as FBCNN and SCUNet

features:

• Remove compression artifacts
• Denoise
• Before/after view
• Fully offline
• Image descaling

models:

You can download models here

examples

Check out examples to get an idea of what DeJPEG can be used for

limitations:

• Processed locally, a fast device is recommended
• Only standard image formats supported

desktop support

chaiNNer is a cross-platform image/model utility, which should work well with these models

For FBCNN, which chaiNNer does support but in a limited fashion, install this custom node and use the original PyTorch models, not the mobile onnx.

building

requirements

• Android SDK
• For full builds: Android NDK (tested on version 27.3.x)
• Optional: UPX for library compression (must be in path if enabled)

steps

1. Clone repo

2. (optional) run ./generatelibs.sh to build native libraries (not needed for lite builds)

   options:

   • --abi <abi>: Target ABI (arm64-v8a, armeabi-v7a, x86_64, x86, or all). Default: arm64-v8a
   • --debug: Build debug variant (disables compression).
   • --no-upx: Disable UPX compression.
   • --full: Build with OpenCV for BRISQUE descaling (requires NDK).
   • --no-cleanup: Reuse existing libraries without rebuilding.
   • --help: Show usage.

   Note: UPX will only work if found in path.

3. ./gradlew clean assembleLiteDebug

   Or open in Android Studio and build from there.

credits and license

disclaimer:

I am by no means a professional developer and only do this in my spare time, the code is not perfect and quite janky.

This is a GUI for a select amount of 1x ONNX processing models, used under their respective licenses (Apache 2.0)

You are welcome to embed parts of this app in your own project as long as it remains free as in beer and abides to the GPLv3 license.

Credits to @adrianerrea for a starting point, FBCNN and SCUNet creators plus all other model owners.

DeJPEG is not affiliated or related with Topaz DEJPEG or any other similarly named software/project. Although I've wondered if the term 'JPEG' is copyrighted/trademarked due to it literally being the acronym for Joint Photographic Experts Group, for this reason I might need to change the app's name if legal issues start to occur.