VQGAN

This project is for personal learning and experimentation, containing implementations of VQGAN.

🔗 The code is primarily adapted from dome272's VQGAN-pytorch repository.

🎨 The dataset used is from This Anime Does Not Exist, which provides high-quality AI-generated anime portraits. These synthetic images are used for training and evaluation purposes in this project.

Requirements

• Python (any version)
• PyTorch (any version)
• CUDA enabled computing device

Results

📊 Below is a comparison between the performance of our model and that of the original repository. Although I used the exact same hyperparameters, my results are noticeably worse. I believe the primary reason is the quality of the dataset.

PS: The dataset I downloaded is actually quite noisy—but I really just wanted an anime dataset, haha.

First Stage (Reconstruction):

50 Epoch(Dome272's VQGAN):

46 Epoch(My VQGAN):

Second Stage (Generating new Images):

Original Left | Reconstruction Middle Left | Completion Middle Right | New Image Right

100 Epoch(Dome272's VQGAN):

20 Epoch(My VQGAN):

Train VQGAN:

First Stage

python train_vqgan.py

Second Stage

python train_transformer_vqgan.py

Citation

@misc{esser2021taming,
      title={Taming Transformers for High-Resolution Image Synthesis}, 
      author={Patrick Esser and Robin Rombach and Björn Ommer},
      year={2021},
      eprint={2012.09841},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}