traiNNer-redux

Overview

traiNNer-redux is a deep learning training framework for image super resolution and restoration which allows you to train PyTorch models for upscaling and restoring images and videos. NVIDIA graphics card is recommended, but AMD works on Linux machines with ROCm.

Usage Instructions

Please see the getting started page for instructions on how to use traiNNer-redux.

Contributing

Please see the contributing page for more info on how to contribute.

Resources

• OpenModelDB: Repository of AI upscaling models, which can be used as pretrain models to train new models. Models trained with this repo can be submitted to OMDB.
• chaiNNer: General purpose tool for AI upscaling and image processing, models trained with this repo can be run on chaiNNer. chaiNNer can also assist with dataset preparation.
• WTP Dataset Destroyer: Tool to degrade high quality images, which can be used to prepare the low quality images for the training dataset.
• Video Destroyer: A tool designed to create datasets from a single high quality input video for AI training purposes.
• helpful-scripts: Collection of scripts written to improve experience training AI models.
• Dataset_Preprocessing: A small collection of scripts for initial dataset processing.
• Enhance Everything! Discord Server: Get help training a model, share upscaling results, submit your trained models, and more.
• vs_align: Video Alignment and Synchonization for Vapoursynth, tool to align LR and HR datasets.
• ImgAlign: Tool for auto aligning, cropping, and scaling HR and LR images for training image based neural networks.
• Deep Learning Tuning Playbook: Document to help you train deep learning models more effectively.

License and Acknowledgement

traiNNer-redux is released under the Apache License 2.0. See LICENSE for individual licenses and acknowledgements.

• This repository is a fork of joeyballentine/traiNNer-redux which itself is a fork of BasicSR.
• Network architectures are imported from Spandrel.
• Several architectures are developed by umzi2: ArtCNN-PyTorch, DUnet, FlexNet, GaterV3, MetaGan, MoESR, MoSR, RTMoSR, SPANPlus
• The ArtCNN architecture is originally developed by Artoriuz.
• The FDAT architecture is created by stinkybread (sharekhan).
• The TFDAT architecture is a temporally stable version of FDAT, created by Kim2091.
• The TSPAN and TSPANv2 architectures are temporally consistent variants of the SPAN architecture created by Kim2091. TSPANv2 is an enhanced version of TSPAN with major improvements to temporal stability, Use Vapourkit to upscale videos with TSPAN and TSPANv2 models.
• The DIS architecture is an ultra lightweight arch created by Kim2091.
• The TSCUNet architecture is from aaf6aa/SCUNet which is a modification of SCUNet, and parts of the training code for TSCUNet are adapted from TSCUNet_Trainer.
• Several enhancements reference implementations from Corpsecreate/neosr and its original repo neosr.
• Members of the Enhance Everything Discord server: Corpsecreate, joeyballentine, Kim2091.