AvatarGAN

A conditional Generative Adversarial Network (cGAN) that generates cartoon avatars conditioned on discrete facial attributes. Built on the CartoonSet dataset.

---

Overview

AvatarGAN learns to generate 128×128 RGB cartoon faces that match a given combination of facial attributes. Both the Generator and the Discriminator receive attribute embeddings alongside the image/noise input, which guides the model towards attribute-consistent outputs.

Architecture

Component	Description
AttributeBlock	Per-attribute two-layer MLP that refines each embedding before it is concatenated
Generator	Latent vector z + attribute embeddings → fully-connected layers → 128×128 RGB image (Tanh output)
Discriminator	Flattened image + L2-normalised attribute embeddings → fully-connected layers → real/fake score (Sigmoid)

Attributes

Four attributes are used, each encoded as a learned embedding:

Attribute	Variants
facial_hair	15
hair	111
face_color	11
hair_color	10

Hyperparameters

Parameter	Value
Image size	128 × 128
Latent dimension	128
Embedding dimension	64
Batch size	64
Epochs	360
Generator LR	0.0002
Discriminator LR	0.0001
Optimizer	Adam (β₁=0.5, β₂=0.999)
Loss	Binary Cross-Entropy

Dataset

CartoonSet10k — 10,000 cartoon avatar images, each paired with a CSV file describing its visual attributes. Place the dataset at ../cartoonset10k relative to the notebook.

cartoonset10k/
├── 00000.png
├── 00000.csv
├── 00001.png
├── 00001.csv
└── ...

Requirements

torch
torchvision
Pillow
matplotlib

Install with:

pip install torch torchvision Pillow matplotlib

Usage

Open and run avatargan.ipynb in Jupyter. The notebook is organised into the following sections:

1. Imports and Configuration — libraries and global hyperparameters
2. Dataset — CustomImageDataset with CSV attribute loading
3. Model Architecture — AttributeBlock, Generator, Discriminator
4. Dataset and Model Setup — dataloader, model initialisation, optimizers
5. Helper Functions — checkpoint saving, fixed sample preparation
6. Training — main training loop with per-epoch visualisation

Checkpoints are saved every 20 epochs to the models4/ directory.

Training Progress

Every epoch a side-by-side comparison of 6 fixed original images and their generated counterparts is displayed, allowing visual tracking of generation quality over time.