SurfaceNet

SurfaceNet is a deep learning framework for estimating spatially-varying BRDFs (SVBRDFs) — including diffuse, normals, roughness, and specular maps — from a single image. Designed and implemented by Zayn Rekhi, the project uses PyTorch and a GAN-style architecture to produce realistic material reconstructions.

---

🔧 Features

• Predicts full SVBRDF maps from a single RGB input
• Adversarial loss and patch-based training for high-fidelity results
• Modular, extensible PyTorch codebase
• Includes training and evaluation scripts
• Supports Accelerate for multi-GPU and distributed training

---

🚀 Getting Started

1. Clone the repository

git clone https://github.com/Zayn-Rekhi/SurfaceNet.git
cd SurfaceNet/src

2. Install dependencies

pip install -r requirements.txt

Make sure you have accelerate configured. If not:

accelerate config

---

🏋️‍♂️ Training

accelerate launch train.py   --tag run_name   --dataset /path/to/data   --logdir logs/

Options:

• --tag: A name for the run
• --dataset: Path to the training dataset
• --logdir: Directory to store logs and checkpoints

---

🧪 Inference / Evaluation

python eval.py   --ckpt path/to/checkpoint.ckpt   --input path/to/image.png   --size 256

The script will output predicted:

• Diffuse map
• Normal map
• Roughness map
• Specular map

---

📁 Directory Structure

SurfaceNet/
├── src/
│   ├── train.py
│   ├── eval.py
│   ├── models/
│   ├── utils/
│   └── ...
├── assets/
├── data/
└── README.md

---

📝 Notes

• Optimized for synthetic BRDF datasets, but can generalize to real-world data with proper preprocessing.
• Outputs are saved in .png format by default.
• Evaluation works on single images; batch inference is easy to enable.

---

🖼 Example Results

(Coming soon – include side-by-side input and SVBRDF outputs here)

---

📄 License

This project is licensed under the MIT License.
Feel free to use, modify, and contribute.

---

🙌 Acknowledgments

Thanks to the open-source PyTorch and GAN communities for foundational tools and techniques.