Implicit Neural Representations (INR) for Super-Resolution and 3D Occupancy

This repository contains an experimental framework for Implicit Neural Representations (INRs): coordinate-based MLPs that represent signals as continuous functions rather than discrete grids.

It supports two pipelines:

• Part 1 — Image Super-Resolution: learn a continuous image function from a low-resolution image and query it at arbitrary resolution.
• Part 2 — 3D Occupancy Reconstruction: learn a continuous occupancy function f(x, y, z) -> {0, 1} from sampled 3D points.

---

Repository overview

Top-level entry points:

• main.py — 2D Super-Resolution experiments
• download_div2k.py — helper to download/prepare DIV2K (if you use DIV2K)
• generate_data.py — generate a 3D occupancy dataset (.npz) from an .obj mesh
• train_3d.py, train_3dv2.py — 3D occupancy training scripts

Core code:

• src/models.py — INR architectures (SIREN / WIRE / Fourier features / MFN / etc.)
• src/config.py — model-specific hyperparameters
• src/trainer.py — training loop, scheduling, batching
• src/utils.py — I/O and visualization utilities

Included sample assets:

• 0788.png — sample image for quick SR sanity-check
• dragon.obj, nefertiti.obj — sample meshes
• dragon_dataset.npz, nefertiti_dataset.npz — pre-generated occupancy datasets

---

Installation

1) Clone

git clone https://github.com/AhmedKElshahed/INR.git
cd INR

2) Create a virtual environment

python -m venv .venv
source .venv/bin/activate      # Linux / macOS
# .venv\Scripts\activate       # Windows PowerShell

3) Install repository dependencies

pip install -r requirements.txt

PART 1 — Image Super-Resolution

Example: Run Super-Resolution

You can specify the input image, number of training epochs, and multiple upscale factors:

python main.py --image 0788.png --epochs 100 --scales 2 4 8 16

PART 2 — 3D Occupancy Reconstruction

Step 1 — Generate Occupancy Dataset from a Mesh

To generate a training dataset from a mesh (e.g., nefertiti.obj), run:

python generate_data.py --mesh nefertiti.obj

Step 2 — Train the Occupancy INR Model

Once the dataset is generated, you can train the 3D INR model using:

python train_3dv2.py --mesh nefertiti.obj --epochs 50