gpt-face-celeb-generator

🧠 Face Generation GPT-2 (VQ-VAE Latent Space)

This model is a GPT-2 fine-tuned to act as an autoregressive generator for human faces. Instead of predicting text tokens, it predicts visual tokens extracted from the CelebA dataset via a VQ-VAE.

🚀 Project Overview

The core idea is to treat image generation as a language modeling problem.

1. Compression: A VQ-VAE encodes $128 \times 128$ face images into $16 \times 16$ discrete tokens (total 256 tokens).
2. Generative Modeling: This GPT-2 model is trained to predict the sequence of these visual tokens based on facial attributes.
3. Reconstruction: The predicted tokens are passed back through the VQ-VAE Decoder to synthesize a new face.

🛠️ Intended Uses & Limitations

How to Use

This model is designed to work in conjunction with the project's custom VQ-VAE decoder. It accepts a sequence of facial attribute tokens and generates the corresponding latent representation of a face.

Limitations

• Resolution: Fixed at $128 \times 128$ reconstruction.
• Realism: While structurally sound, it may lack the high-frequency details found in modern Diffusion models.
• Dependencies: Requires the specific VQ-VAE Codebook and Decoder available in the GitHub Repository.

📊 Training Data

• Dataset: CelebA (CelebFaces Attributes Dataset).
• Preprocessing: Images resized to $128 \times 128$, normalized, and tokenized using a Vector Quantized Variational Autoencoder.
• Input Format: <START_FACE> <ATTRIBUTES> <START_GENERATION> <256_TOKENS> <END_GENERATION>.

⚙️ Training Procedure

Hyperparameters

• Learning Rate: 2e-4
• Batch Size: 40 (Effective Batch Size: 80 with 2 Gradient Accumulation steps)
• Precision: Mixed Precision (Native AMP)
• Epochs: 5
• Optimizer: AdamW (Fused)

Training Results

The model shows steady convergence in cross-entropy loss over 5 epochs:

🔗 Resources

• Live Demo: Streamlit Web App
• Source Code: GitHub Repository

📜 Framework Versions

• Transformers 5.0.0
• Pytorch 2.10.0+cu128
• Datasets 4.8.3
• Tokenizers 0.22.2