bilal3406/DDPM-Face-Generator

DDPM Face Generator & Reconstructor (From Scratch)

This Space hosts a Denoising Diffusion Probabilistic Model (DDPM) built entirely from scratch using base PyTorch, without relying on abstracted pipelines like HuggingFace Diffusers. It was trained on a subset of the FFHQ (Flickr-Faces-HQ) dataset.

🏗️ Architecture Details

• —Framework: Pure PyTorch.
• —Backbone: Custom U-Net with a channel progression of 64 → 128 → 256.
• —Time Awareness: Sinusoidal Positional Embeddings injected into every Residual Block.
• —Global Structure: A Self-Attention block deployed at the $32 \times 32$ bottleneck to help the model learn spatial symmetry.
• —Forward Process: Linear variance schedule over $T = 300$ timesteps.

⚠️ Training & Hardware Constraints

This model was engineered to train within the strict hardware limits of a Kaggle Notebook (Dual T4 GPUs with 16GB VRAM, and a 30GB System RAM limit).

• —To prevent VRAM fragmentation, it utilizes Automatic Mixed Precision (AMP) and aggressive garbage collection.
• —To prevent System RAM leaks, the PyTorch DataLoader is restricted to the main thread.
• —Note on Visual Quality: Due to compute time limitations, these weights were trained for ~20 epochs. Therefore, the model has successfully learned local distributions (skin tones, hair textures) but lacks the compute time required to fully lock in global facial symmetry (resulting in the current abstract, "Picasso-style" outputs).

🚀 How to Use This Space

This interactive Gradio application features two main tabs:

1. 1.Generate Face: Performs unconditional generation. The model starts with a tensor of pure Gaussian noise and iteratively removes the noise over 300 steps to hallucinate a novel face. No input required.
2. 2.Reconstruct & Evaluate: Upload a target image. The script will apply maximum noise to destroy the image, and then force the U-Net to walk backward to reconstruct it. The app calculates PSNR and SSIM scores to quantitatively evaluate the spatial accuracy of the reconstruction.

👥 Authors

Engineered and trained by:

• —Muhammad Bilal Ahmad Ali Hamza FAST National University of Computer and Emerging Sciences (FAST NUCES)