... is a free Android app for removing jpeg compression artifacts and denoising images using FBCNN and SCUNet.
- batch processing
- transparancy re-mapping (experimental)
- fully offline
note: this is not a "super resolution AI upscaler", but a simple non-destructive way to clean up compressed/noisy images
FBCNN (JPEG artifacts)
SCUNet (Grain/Noise)
info about models here (also in the apps FAQ)
- processed on your phone (no hardware acceleration), I'd recommend a device with ~4gb of ram and a CPU with at least 4 threads
DeJPEG is not affiliated or related to Topaz's DEJPEG or any other software/project
for Qualcomm devs
tl;dr: Snapdragon devices support a special type of hardware acceleration, but these models wouldn't benefit from it.
I've looked into Hexagon/HTP support for accelerating the speed of the processing but the models from my internal testing perform better with CPU.
which means if I were to use QCOMMs special HWA, the model would still need to use the CPU for some parts of processing.
best I can understand it as HTP doesn't support some of FBCNNs operations and would shift back and forth from HTP to CPU during any image, which isn't really beneficial.
I'd also need a Qualcomm developer license requiring me to hand over a lot of personal information for simply what I consider a hobby.
and retraining/recreating the models aside from simple conversion, is outside my expertise.
@adrianerrea for information on how to create a mobile PyTorch app and various information which made it easier to run the ONNX model
do you use AI?
I partially use Generative AI, (ChatGPT, Github Copilot, et. al.) for some complicated tasks and problems.
AI has its place, as a tool, not as a replacement.
I only use it in my IDE, all my comments, changelogs, and interactions with me are typed with my hands, from my brain.
I do not pay for access to generative AI as well, i use whatever free tier is available
If you see Google's "Jules AI" as a contributor (if its still there), I tried it out and it created a PR based on things I described, but didn't go as planned and reverted all changes.
and no, this text wasn't written by AI.
all models used are under their respective licenses
you are free to embed parts of this app in your own project as long as it remains free/non-paywalled and must abide to the GPL v3 license