https://github.com/soroushesnaashari/anime-faces-generator-with-gan-using-tensorflow

A GAN project using "TensorFlow" creating model to generate new random Anime faces
https://github.com/soroushesnaashari/anime-faces-generator-with-gan-using-tensorflow

deep-learning gan neural-network tensorflow

Last synced: 9 months ago
JSON representation

A GAN project using "TensorFlow" creating model to generate new random Anime faces

• Host: GitHub
• URL: https://github.com/soroushesnaashari/anime-faces-generator-with-gan-using-tensorflow
• Owner: soroushesnaashari
• Created: 2025-05-21T07:21:21.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2025-06-12T07:12:56.000Z (about 1 year ago)
• Last Synced: 2025-06-21T22:36:45.896Z (about 1 year ago)
• Topics: deep-learning, gan, neural-network, tensorflow
• Language: Jupyter Notebook
• Homepage:
• Size: 15.7 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
## Anime Faces Generator with GAN using TensorFlow

[![](images/Image.gif)](https://www.kaggle.com/code/soroushesnaashari/anime-faces-generator-gan-using-tensorflow/output)

### Overview

This project implements a Deep Convolutional Generative Adversarial Network (DCGAN) in TensorFlow/Keras to generate anime-style faces. By training on a curated dataset of ~21 k anime character portraits (cropped and resized to 64×64 px), the GAN learns to map a 100-dimensional latent vector to realistic-looking anime faces. The entire pipeline, from data ingestion through model definition, training, and image sampling—is contained in this repository.




### Project Flow

1. **Data Preparation**  

   - Download anime face images (21 551 samples) from the “Anime Faces” dataset.  

   - Crop to facial regions and resize all images to 64×64 pixels.  

   - Normalize pixel values to \[-1, +1\] for GAN training.

2. **Model Definition**  

   - **Generator:**  

     - Input: 100-dimensional noise vector.  

     - Upsampling path: series of `Conv2DTranspose` layers with increasing spatial resolution (4→8→16→32→64), each followed by BatchNorm + ReLU, ending in a 64×64×3 `tanh` output.  

   - **Discriminator:**  

     - Input: 64×64×3 image.  

     - Downsampling path: series of `Conv2D` layers with LeakyReLU activations and Dropout, reducing to a single scalar probability via a final dense layer with `sigmoid`.

3. **Training Loop**  

   - Batch size: 256  

   - Latent dimension: 100  

   - Optimizer: Adam (learning rate 2e-4, β₁ = 0.5)  

   - Number of epochs: 50  

   - At each step:  

     1. Sample a batch of real anime faces and a batch of random noise.  

     2. Update the discriminator on real vs. generated (“fake”) images.  

     3. Update the generator via its ability to fool the discriminator.  

   - Periodically save checkpointed model weights and sample grids of generated faces.

4. **Sampling & Visualization**  

   - After every N epochs (e.g. 5), generate a 5×5 grid of new faces from fixed noise vectors.  

   - Plot and save loss curves for both generator and discriminator.




### Key Features

- **Custom DCGAN Architecture**  

  Lightweight yet expressive convolutional generator and discriminator networks tailored for 64×64 anime faces.

- **TensorFlow 2.x / Keras**  

  Leverages the `tf.GradientTape` API for full control over training; easily extensible.

- **Flexible Hyperparameters**  

  Easily adjust latent dimension, batch size, learning rates, and number of epochs via the configuration section.

- **Built-in Visualization**  

  Automatic generation of sample image grids and loss plots to monitor training progress.

- **Checkpoints & Resume**  

  Save and load model weights to resume long-running training jobs.




### Results

After 50 epochs of training:

- The generator produces high-quality anime faces with coherent hair styles, facial features, and coloring.

- Sample outputs (5×5 grid at epochs 5, 15, 30):  

  ![Samples @ Ep 5](images/epoch05.png)  

  ![Samples @ Ep 15](images/epoch15.png)  

  ![Samples @ Ep 30](images/epoch30.png)

- After 50 epoches the model generated some acceptable anime characters (bottom image) which can be improved a lot with some more profeesional GAN model and also with more epoches.

  ![Fresh image](images/fresh.png)




### Repository Contents

- **`anime-faces-generator-gan-using-tensorflow.ipynb`**: Jupyter Notebook with full code, visualizations and explanations.

- **`Data`:** Contains the [Original Dataset](https://www.kaggle.com/datasets/splcher/animefacedataset) and you can see the cleaned dataset in notebook.

- **`README.md`:** Project documentation.




### How to Contribute

Contributions are welcome! If you'd like to improve the project or add new features:

1. **Fork the repository.**

2. **Create a new branch.**

3. **Make your changes and submit a pull request.**