Pixel

A PyTorch-based Generative Adversarial Network (GAN) for training and generating pixel art images.

Table of Contents

• Overview
• Project Structure
• Architecture
• Workflow
• Installation
• User Guide
• Configuration
• Examples

---

Overview

This project implements a Deep Convolutional GAN (DCGAN) to generate pixel art images. It includes:

• Training Pipeline: Train a GAN model on your custom image dataset
• Image Generation: Generate new images using trained models
• Interactive GUI: Tkinter-based interface for real-time generation
• GGUF Support: Convert and use GGUF quantized models

---

Project Structure

pixel-model-trainer/
│
├── trainer.py                  # GAN training script
├── generator.py                # Image generation (GUI + CLI)
│
├── data/                       # Training data
│   ├── attributes.csv          # Dataset metadata
│   └── images/                 # Training images (punk000.png, punk001.png, ...)
│
├── models/                     # Trained model storage
│   └── generator_model.safetensors
│
├── generated/                  # Generated image outputs
│   ├── output.png              # Grid visualizations
│   └── individual/             # Individual generated images
│
├── gen_images/                 # Training progress images
│   ├── epoch_0.png
│   ├── epoch_1.png
│   └── ...
│
└── gguf/                       # GGUF format support
    └── generator.py            # GGUF model converter/generator

Directory Purpose

---

Architecture

GAN Components

┌─────────────────────────────────────────────────────────────┐
│                      GAN Architecture                       │
└─────────────────────────────────────────────────────────────┘

┌──────────────────┐                    ┌──────────────────┐
│   Generator      │                    │  Discriminator   │
│                  │                    │                  │
│  Input: Noise    │                    │  Input: Images   │
│  (100-dim)       │                    │  (24x24x4)       │
│                  │                    │                  │
│  ┌────────────┐  │                    │  ┌────────────┐  │
│  │   FC       │  │                    │  │   Conv2D   │  │
│  │  (9,216)   │  │                    │  │  64 filters│  │
│  └────────────┘  │                    │  └────────────┘  │
│        ↓         │                    │        ↓         │
│  ┌────────────┐  │                    │  ┌────────────┐  │
│  │ Reshape    │  │    ┌──────────┐    │  │   Conv2D   │  │
│  │ (256,6,6)  │  │───→│  Real or │←───│  │ 128 filters│  │
│  └────────────┘  │    │   Fake?  │    │  └────────────┘  │
│        ↓         │    └──────────┘    │        ↓         │
│  ┌────────────┐  │                    │  ┌────────────┐  │
│  │ ConvTrans  │  │                    │  │   Conv2D   │  │
│  │ 128 filters│  │                    │  │ 256 filters│  │
│  └────────────┘  │                    │  └────────────┘  │
│        ↓         │                    │        ↓         │
│  ┌────────────┐  │                    │  ┌────────────┐  │
│  │ ConvTrans  │  │                    │  │  GlobalAvg │  │
│  │  64 filters│  │                    │  │    Pool    │  │
│  └────────────┘  │                    │  └────────────┘  │
│        ↓         │                    │        ↓         │
│  ┌────────────┐  │                    │  ┌────────────┐  │
│  │ ConvTrans  │  │                    │  │  FC + Sig  │  │
│  │  4 channels│  │                    │  │  (0-1)     │  │
│  └────────────┘  │                    │  └────────────┘  │
│                  │                    │                  │
│  Output: Image   │                    │  Output: Score   │
│  (24x24x4 RGBA)  │                    │  (Real/Fake)     │
└──────────────────┘                    └──────────────────┘

Model Details

Generator:

• Input: 100-dimensional latent vector (random noise)
• Architecture: FC → BatchNorm → 3x ConvTranspose2D with BatchNorm
• Output: 24x24x4 RGBA image (values in [-1, 1])
• Activation: LeakyReLU + Tanh (output)

Discriminator:

• Input: 24x24x4 RGBA image
• Architecture: 3x Conv2D with Dropout → GlobalAvgPool → FC
• Output: Probability score [0, 1] (real vs. fake)
• Activation: LeakyReLU + Sigmoid (output)

---

Workflow

Training Workflow

┌─────────────┐
│  Start      │
└──────┬──────┘
       │
       ▼
┌──────────────────────┐
│  Load Dataset        │
│  (data/images/)      │
└──────┬───────────────┘
       │
       ▼
┌──────────────────────┐
│  Initialize Models   │
│  - Generator         │
│  - Discriminator     │
└──────┬───────────────┘
       │
       ▼
┌──────────────────────┐
│  Training Loop       │◄─────────┐
│  (N epochs)          │          │
└──────┬───────────────┘          │
       │                          │
       ▼                          │
┌──────────────────────┐          │
│  For each batch:     │          │
│  1. Train Discrim.   │          │
│  2. Train Generator  │          │
└──────┬───────────────┘          │
       │                          │
       ▼                          │
┌──────────────────────┐          │
│  Save Progress       │          │
│  (gen_images/)       │──────────┘
└──────┬───────────────┘
       │
       ▼
┌──────────────────────┐
│  Save Final Model    │
│ (models/*.safetensors)
└──────┬───────────────┘
       │
       ▼
┌──────────────┐
│  Complete    │
└──────────────┘

Generation Workflow

┌─────────────────────┐
│  Mode Selection     │
└──────┬──────────────┘
       │
       ├─────────────────────────┐
       │                         │
       ▼                         ▼
┌──────────────┐       ┌──────────────────┐
│  GUI Mode    │       │   CLI Mode       │
└──────┬───────┘       └──────┬───────────┘
       │                      │
       ▼                      ▼
┌──────────────┐       ┌──────────────────┐
│ Load Model   │       │ Load Model       │
│ (safetensors)│       │ Parse Args       │
└──────┬───────┘       └──────┬───────────┘
       │                      │
       ▼                      ▼
┌──────────────┐       ┌──────────────────┐
│ Tkinter GUI  │       │ Generate N Images│
│ - Button 1x1 │       │ - Custom grid    │
│ - Button 3x3 │       │ - Custom seed    │
│ - Button 5x5 │       │ - Save options   │
└──────┬───────┘       └──────┬───────────┘
       │                      │
       ▼                      ▼
┌──────────────┐       ┌──────────────────┐
│ On Click:    │       │ Save Grid        │
│ Generate     │       │ Save Individual  │
│ Display      │       │ (optional)       │
└──────┬───────┘       └──────────────────┘
       │
       ▼
┌──────────────┐
│ Interactive  │
│ Generation   │
└──────────────┘

---

Installation

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (optional, for faster training)

Setup

1. Clone/Download the repository

2. Install dependencies:

pip install torch torchvision
pip install numpy pandas matplotlib pillow
pip install safetensors

3. Prepare your dataset:

Place your training images in data/images/ with filenames like punk000.png, punk001.png, etc.

Create data/attributes.csv:

id
0
1
2
...

*this is a community dataset, serves as an example or demo, we do not tier to crytopunk project, you could replace it with your own dataset (see details under Examples session)

---

User Guide

1. Training a Model

Train the GAN on your dataset:

python trainer.py \
    --data_path ./data/attributes.csv \
    --images_path ./data/images/ \
    --model_output_path ./models/ \
    --images_output_path ./gen_images/ \
    --epochs 50 \
    --batch_size 16

Training Parameters:

Training Output:

• Progress displayed: Epoch X/Y - Gen Loss: X.XXXX, Disc Loss: X.XXXX
• Progress images saved to gen_images/epoch_N.png
• Final model saved to models/generator_model.safetensors

---

2. Generating Images

Option A: Interactive GUI Mode (Default)

Launch the Tkinter GUI for real-time generation:

python generator.py

or explicitly:

python generator.py --gui

GUI Controls:

• Generate 1 avatar: Single image
• Generate 3x3 avatars: 3x3 grid (9 images)
• Generate 5x5 avatars: 5x5 grid (25 images)
• Terminate: Close the application

---

Option B: Command-Line Interface (CLI)

Batch generate images from terminal:

Basic generation (16 images):

python generator.py --num_images 16 --output_path ./generated/output.png

Custom grid (8x8 = 64 images):

python generator.py --grid_size 8 --output_path ./generated/grid_8x8.png

Reproducible generation (with seed):

python generator.py --grid_size 4 --seed 42 --output_path ./generated/seed42.png

Save individual images:

python generator.py \
    --num_images 100 \
    --save_individual \
    --individual_output_dir ./generated/individual/ \
    --output_path ./generated/batch.png

CLI Parameters:

---

3. GGUF Model Support

Use quantized GGUF models for smaller file sizes:

cd gguf/
python generator.py

The GGUF generator will:

1. Detect available .gguf files in the directory
2. Prompt you to select a model
3. Convert GGUF → SafeTensors format
4. Launch the standard generator

---

Configuration

Model Architecture Configuration

Modify these parameters in both trainer.py and generator.py:

--codings_size 100        # Latent vector dimension
--image_size 24           # Output image size
--image_channels 4        # RGBA (4) or RGB (3)

Training Hyperparameters

In trainer.py:

# Optimizer
gen_optimizer = optim.RMSprop(generator.parameters(), lr=0.001)
disc_optimizer = optim.RMSprop(discriminator.parameters(), lr=0.001)

# Loss function
criterion = nn.BCELoss()

# Dropout rate (in Discriminator)
nn.Dropout(0.4)

Data Preprocessing

In trainer.py → ImageDataset:

transforms.Compose([
    transforms.Resize((image_size, image_size)),
    transforms.ToTensor(),
    transforms.Normalize([0.5] * channels, [0.5] * channels)  # [-1, 1]
])

---

Examples

Example 1: Train on Custom Dataset

# Prepare your data
# data/images/punk000.png, punk001.png, ..., punk099.png
# data/attributes.csv with ids 0-99

# Train for 100 epochs
python trainer.py \
    --data_path ./data/attributes.csv \
    --images_path ./data/images/ \
    --epochs 100 \
    --batch_size 32 \
    --model_output_path ./models/my_model.safetensors

Example 2: Generate with Specific Seed

# Generate same images every time
python generator.py \
    --model_path ./models/generator_model.safetensors \
    --grid_size 5 \
    --seed 12345 \
    --output_path ./results/reproducible.png

Example 3: Batch Generation

# Generate 1000 individual images
python generator.py \
    --num_images 1000 \
    --save_individual \
    --individual_output_dir ./dataset_synthetic/ \
    --output_path ./dataset_synthetic/overview.png

Example 4: Monitor Training Progress

# Training with progress visualization
python trainer.py \
    --epochs 200 \
    --images_output_path ./training_progress/

# View progress images
ls ./training_progress/
# epoch_0.png, epoch_1.png, ..., epoch_199.png

---

Technical Details

Model File Format

Models are saved in SafeTensors format (.safetensors) with embedded metadata:

metadata = {
    'codings_size': '100',
    'image_size': '24',
    'image_channels': '4'
}

This ensures the generator automatically loads the correct architecture.

Image Value Ranges

• Training: Images normalized to [-1, 1]
• Generation output: Images scaled to [0, 1]
• Saved files: Images saved as uint8 [0, 255]

GPU Support

The code automatically detects and uses CUDA if available:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

---

Troubleshooting

Q: Training loss not decreasing?

• Try adjusting learning rates
• Increase batch size or epochs
• Check if dataset has sufficient variety

Q: Generated images look like noise?

• Model needs more training epochs
• Dataset may be too small (need 50+ images minimum)
• Try adjusting discriminator dropout rate

Q: GUI not launching?

• Check Tkinter installation: python -m tkinter
• On Linux: sudo apt-get install python3-tk

Q: CUDA out of memory?

• Reduce batch size: --batch_size 8
• Reduce image size: --image_size 16

---

License

This project is provided as-is for educational and creative purposes.

---

Acknowledgments

• Built with PyTorch
• Inspired by DCGAN architecture
• Uses SafeTensors for model serialization