AdGen

An Auto-Generator for Ads with Never-Seen-Before Humans

More about the project: Medium Blog, Addendum, Slides

• You will find all the dataset files in npy format at this Google Drive Link: dataset and the web crawled and pre-processed images in the CrawlingImages folder of this repo.

• To run the StyleGAN2 model: Simply run the test.sh file available in the stylegan2 folder to generate images of brand new never-seen-before humans.

• To run the GFLA model:

  • Clone the repository from the original paper found here: GFLA repo and then run the test.py file to generate the human models in the required poses.
  • Run createPairsCSV.py to create the pairs that map the output of GFLA to output of StyleGAN2.
  • The output of the StyleGAN2 + GFLA method can be found in the fashion_900.zip file in the dataset Google Drive.

• To run the Disentanglement model:

  • Clone the repository from the original paper found here: Disentangled Person Image Generation Method repo and then run the run_DF_test.sh script to generate brand new humans in fixed poses sampled from noise and trained on DeepFashion Dataset.
  • You will also need to modify the code in run_DF_test.sh and tester.py to sample appearances and keep the poses fixed which is different from the default available in the files.
  • The 'G' folder in the dataset Google Drive link has the output of the Disentanglement method.

• To calculate the MSE Loss or Inception scores of the StyleGAN2 + GFLA and Disentanglement approaches:

  • You can use the files available inside comparisonOfModels folder of this repo.
  • To calculate MSE Loss between input and output poses, run mselosscalc.py and replace the estimator.py file with the estimator.py file available in the repo.

• To run the ST-GAN model on the DeepFashion dataset:

  • Please find all the training and test files in the st-gan folder and simply download the code along with the datasets available in the dataset google drive link and place the datasets in the my_data folder inside the st-gan folder.
  • Make sure to change the local paths in data.py and in the test and train scripts if you are running on Windows.