Comparing the visual quality of deep generative models for steel microstructures

This repository provides the PyTorch code used in our expert study to compare the perceived realism of synthetic steel microstructures of two deep generative models:

@article{wentzien2025visualqualitysteelmicrostructures,
author = {Wentzien, Marcel and Ingber, Jerome and Schloetterer, Joerg and Schmalzried, Dirk},
title = {Comparing the visual quality of deep generative models for steel microstructures},
year = {2025},
journal = {TBA},
doi = {TBA},
}

Visual Abstract

(FractalGen image displayed as in the original paper.)

Deep generated models

Original FractalGen repository: https://github.com/LTH14/fractalgen

Original StyleGANv2-Ada code is part of the StyleGANv3 repository: https://github.com/NVlabs/stylegan3

Training a model

The following code assumes you have working installations of the original repositories. They were tested as SLURM script on NVIDIA A100 GPUs.

FractalGen

Example script for training 800 epochs on 256x256 images of 30 different steel alloys.

OUTPUT=/home/models/fractalgen/
DATA=/home/data/

torchrun --nproc_per_node=2 --nnodes=1 --node_rank=0 --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} \
main_fractalgen.py \
--model fractalmar_base_in256 --img_size 256 --num_conds 5 --guiding_pixel \
--batch_size 32 --eval_freq 80 --save_last_freq 10 --class_num 30 \
--epochs 800 --warmup_epochs 1 \
--blr 5.0e-4 --weight_decay 0.005 --r_weight 1. --attn_dropout 0.1 --proj_dropout 0.1 --lr_schedule cosine \
--gen_bsz 256 --num_images 1272 --num_iter_list 64,16,16 --cfg 10.0 --cfg_schedule linear --temperature 1.1 \
--output_dir $OUTPUT \
--data_path $DATA --num_worker 4 --grad_checkpointing --online_eval

StyleGANv2-Ada

OUTPUT=/home/models/stylegan/
DATA=/home/data/

python train.py --outdir=$OUTPUT --cfg=stylegan2 --data=$DATA --batch=32 --gpus=2 --gamma=8 --target=0.5 --aug=ada --kimg=5000 --tick=2 --snap=30 --nobench=True --workers=2 --metrics=kid50k_full --mirror=1

Generating new microstructure images

FractalGen

Example script for generating new microstructural images based on FractalGen's demo. Ensure to put generate_fractalgen.py in the original FractalGen repository as it depends on the original code. We found it best to run the script on a SLURM cluster. On SLURM cluster run from within a cloned FractalGen repository:

torchrun --nproc_per_node=1 --nnodes=1 --node_rank=0 --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} \
generate_fractalgen.py 

StyleGANv2-Ada

Ensure a fully trained model pkl-file is located under model/. Code assumes CUDA is available. Then run

python src/generate_stylegan.py

to generate new images under model/generated_fakes/

Classifying the images with the GAN's discriminator

In order to classify images in the data/ folder using the GAN's discriminator, ensure to correctly set up all paths in src/discriminate_images.py. This includes providing a trained StyleGANv2 model in model/.

The class_fake subfolder stubs are only required for simplified dataloading using pytorch's ImageFolder.

Ensure PyTorch and Torchvision are installed, see Pytorch Website

Then run

python src/discriminate_images.py

By default the results will be stored in evaluate/discriminator/*/discriminations.csv where * is the same subfolder as provided under data/.

Classifying the images as human expert using the custom UI

Either use the provided sample data or ensure that your own data is correctly located in the subfolders under data/.

Ensure PyQt6 is installed.

Then run

python src/real_or_fake_ui.py

By default the results will be stored in evaluate/guesses.csv.