Shell scripts to prepare textures for ESR/SFTGAN etc.

Table of Contents

• Description
  • Scripts to prepare for inference
  • Scripts to prepare for training
• Installing ESRGAN and/or SFTGAN
  • If you want to use ESRGAN to upscale
  • If you want to use SFTGAN to upscale
  • If you want to train SFTGAN and/or SFTGAN
• Usage for inference/upscaling with ESRGAN and/or SFTGAN
  • Steps when you want to use ESRGAN only
  • Steps when you want to use SFTGAN only
  • Steps when you want to use ESRGAN and then SFTGAN
  • Steps when you want to use SFTGAN and then ESRGAN
• Usage for training ESRGAN and/or SFTGAN
• Troubleshooting
• Requirements
  • Required if you want to use ESRGAN and/or SFTGAN
  • Required if you want to train ESRGAN

Description

The included scripts are for preparing your images to be used by ESRGAN, SFTGAN etc. and ressassemble them (no seams)

Note: Will only work on power of 2 sized textures at the moment

Note: When . is mentioned before a path, it means relative to the ctp-texture-upscale directory

Scripts to prepare for inference

step1_create_tiles.sh: Create tiles with overlap, separating the RGB and alpha, with optional rescaling

step2_copy_tiles.sh: Copy tiles from one directory to the next, with optional rescaling while copying

step3_assemble_tiles.sh: Reassemble tiles and use the overlap for blending (to remove seams), recombine the RGB and alpha, with optional rescaling

Scripts to prepare for training

training/step1_create_tiles.sh: Create equal size tiles (1 for HR/GT, 1 downscaled for LR), separating the RGB and alpha, use separate directories according to regexp

training/step2_cleanup_tiles.sh: Cleanup tiles, remove tiles that have too little colors and/or that do not fit the required size for HR/GT and LR

training/step3_select_tiles.sh: Select tiles according to a specified percentage to be used for training and validation

Installing ESRGAN and/or SFTGAN

If you want to use ESRGAN to upscale

• Install ESRGAN to ./esrgan
• Follow the instructions for installing ESRGAN at: https://github.com/xinntao/ESRGAN

If you want to use SFTGAN to upscale

• Install SFTGAN to ./sftgan
• Follow the instructions for installing SFTGAN at: https://github.com/xinntao/SFTGAN

If you want to train SFTGAN and/or SFTGAN

• Install BasicSR to ./basicsr
• Follow the instructions for installing BasicSR at: https://github.com/xinntao/BasicSR

Usage for inference/upscaling with ESRGAN and/or SFTGAN

Steps when you want to use ESRGAN only

• Put all the textures you want to process inside the ./input directory
• ./step1_create_tiles.sh --output-dir="./esrgan/LR"
• pushd ./esrgan/; python3 test.py ./models/RRDB_ESRGAN_x4.pth; popd (Replace model path if you want)
• ./step3_assemble_tiles.sh --input-dir="./esrgan/results" --input-postfix="_rlt"
• Results will be inside ./output

Steps when you want to use SFTGAN only

• Put all the textures you want to process inside the ./input directory
• ./step1_create_tiles.sh --output-dir="sftgan/data/samples" --resize="200%" (SFTGAN requires you to upscale first)
• pushd ./sftgan/pytorch_test/; python3 test_segmentation.py; popd
• pushd ./sftgan/pytorch_test/; python3 test_sftgan.py; popd
• ./step3_assemble_tiles.sh --input-dir="./sftgan/data/samples_result" --input-postfix="_rlt"
• Results will be inside ./output

Steps when you want to use ESRGAN and then SFTGAN

• Put all the textures you want to process inside the ./input directory
• ./step1_create_tiles.sh --output-dir="./esrgan/LR"
• pushd ./esrgan/; python3 test.py ./models/RRDB_ESRGAN_x4.pth; popd (Replace model path if you want)
• ./step2_copy_tiles.sh --input-dir="./esrgan/results" --output-dir"./sftgan/data/samples" --input-postfix="_rlt"
• pushd ./sftgan/pytorch_test/; python3 test_segmentation.py; popd
• pushd ./sftgan/pytorch_test/; python3 test_sftgan.py; popd
• ./step3_assemble_tiles.sh --input-dir="./sftgan/data/samples_result" --input-postfix="_rlt"
• Results will be inside ./output

Steps when you want to use SFTGAN and then ESRGAN

• Put all the textures you want to process inside the ./input directory
• ./step1_create_tiles.sh --output-dir="./sftgan/data/samples" --resize="200%" (SFTGAN requires you to upscale first)
• pushd ./sftgan/pytorch_test/; python3 test_segmentation.py; popd
• pushd ./sftgan/pytorch_test/; python3 test_sftgan.py; popd
• ./step2_copy_tiles.sh --input-dir="./sftgan/data/samples_result" --output-dir="./esrgan/LR" --input-postfix="_rlt"
• pushd ./esrgan/; python3 test.py ./models/RRDB_ESRGAN_x4.pth; popd (Replace model path if you want)
• ./step3_assemble_tiles.sh --input-dir="./esrgan/results" --input-postfix="_rlt"
• Results will be inside ./output

Note: Try out different combinations of settings, use ./<script>.sh --help

Note: Upscaling the textures before running SFT may give better results

Note: Everything works much better on Linux

Usage for training ESRGAN and/or SFTGAN

• Put all textures you want to use for training in ./training/input
• Go to the ./training directory
• ./step1_create_tiles.sh
• Results will be inside ./training/output
• ./step2_cleanup_tiles.sh
• ./step3_select_tiles.sh
• End result will be inside ./training/output_training and ./training/output_validation
• Go to <basicsr path>/codes and modify the options/train/train_ESRGAN.json file as follows:
  • Change the name to something else that has no debug in it.
  • In train you may want to decrease the n_workers and batch_size values to for ex.: 4 and 8
  • In train point the dataroot_LR to <path>/training/output_training/LR
  • In train point the dataroot_HR to <path>/training/output_training/HR
  • In val point the dataroot_LR to <path>/training/output_validation/LR
  • In val point the dataroot_HR to <path>/training/output_validation/HR
  • In path change root to the folder where BasicSR is installed
  • In path you may have to remove resume_state for now
• In <basicsr path>/codes run the following: python3 train.py -opt options/train/train_ESRGAN.json
• Results will be inside <basicsr path>/experiments, after every X iterations it will write some validation results
• If you are happy with the validation results you can stop the training using Ctrl+C
• If you want to continue training make sure that in train_ESRGAN.json the resume_state is set in path pointing to the state file <basicsr path>/experiments/<name>/training_state/<state file>.state you want to continue training from

Troubleshooting

• You may have to increase some of the values in Image Magick's policy.xml file to allow for more memory to be used.
• If Image Magick keeps complaining about memory limits you can comment out the resource lines in policy.xml

Requirements

• Bash (if you do not have Linux you can try the Git for Window's Bash instead)
• Image Magick

Required if you want to use ESRGAN and/or SFTGAN

• Cuda (optional, but recommended if you have an NVidia) (On Linux you can just install the latest NVidia drivers)
• ESRGAN: https://github.com/xinntao/ESRGAN
• SFTGAN: https://github.com/xinntao/SFTGAN
• Python 3 64-bit
• PyTorch

Required if you want to train ESRGAN

• Python modules: numpy opencv-python torchvision tensorboardX lmdb
• BasicSR: https://github.com/xinntao/BasicSR/