diff --git a/README.md b/README.md
index f0f50e60..645224d9 100644
--- a/README.md
+++ b/README.md
@@ -115,4 +115,48 @@ Finally, DeepLense help combat the problem of noisy and low-resolution of real l
 **Pranath Reddy** performs a comparative study of the super-resolution of strong lensing images in their [GSoC 2023 project](https://summerofcode.withgoogle.com/archive/2023/projects/Rh8kJLr4), using Residual Models with Content Loss and Conditional Diffusion Models, on the Model 1 dataset.
 
 #### 3.3.3 Physics-Informed Unsupervised Super-Resolution of Strong Lensing Images
-**Anirudh Shankar** explores the unsupervised super-resolution of strong lensing images through a Physics-Informed approach in his [GSoC 2024 project](https://summerofcode.withgoogle.com/programs/2024/projects/AvlaMMJJ), built to handle sparse datasets. They use custom datasets using different lens models and light profiles. 
\ No newline at end of file
+**Anirudh Shankar** explores the unsupervised super-resolution of strong lensing images through a Physics-Informed approach in his [GSoC 2024 project](https://summerofcode.withgoogle.com/programs/2024/projects/AvlaMMJJ), built to handle sparse datasets. They use custom datasets using different lens models and light profiles.
+## 🆕 Loading `.npy` Dataset
+
+This repository now supports loading datasets stored in `.npy` format using a custom PyTorch Dataset class.
+
+### 📌 Usage
+
+```python
+from dataset.npy_loader import NPYDataset
+
+# Initialize dataset
+dataset = NPYDataset(data_dir="path/to/npy_files")
+
+# Get one sample
+image, label = dataset[0]
+
+print(image.shape, label)
+```
+
+### 📂 Supported Formats
+
+* `(H, W)` → Grayscale images
+* `(H, W, C)` → RGB / multi-channel images
+
+### ⚠️ Notes
+
+* Ensure all `.npy` files contain valid NumPy arrays
+* All files should have consistent shape
+* Labels can be:
+
+  * encoded in filenames (e.g., `cat_1.npy`, `dog_2.npy`)
+  * or handled separately in your dataset class
+
+### 💡 Example Directory Structure
+
+```
+data/
+├── cat_1.npy
+├── cat_2.npy
+├── dog_1.npy
+```
+
+### 🚀 Benefit
+
+This feature allows direct training on `.npy` datasets without converting them into image formats, improving efficiency and flexibility.
diff --git a/npy_dataset.py b/npy_dataset.py
new file mode 100644
index 00000000..d3969e8c
--- /dev/null
+++ b/npy_dataset.py
@@ -0,0 +1,50 @@
+import os
+import torch
+from torch.utils.data import Dataset
+import numpy as np
+
+class NPYDataset(Dataset):
+    """
+    Custom PyTorch Dataset for loading .npy-based DeepLense data.
+    """
+
+    def __init__(self, root_dir):
+        self.samples = []
+
+        # Only include directories (ignore files like dataset.py)
+        self.class_names = [
+            d for d in os.listdir(root_dir)
+            if os.path.isdir(os.path.join(root_dir, d))
+        ]
+
+        self.class_to_idx = {cls: i for i, cls in enumerate(self.class_names)}
+
+        for cls in self.class_names:
+            class_dir = os.path.join(root_dir, cls)
+
+            for file in os.listdir(class_dir):
+                if file.endswith(".npy"):
+                    self.samples.append(
+                        (os.path.join(class_dir, file), self.class_to_idx[cls])
+                    )
+
+    def __len__(self):
+        return len(self.samples)
+
+    def __getitem__(self, idx):
+        file_path, label = self.samples[idx]
+
+        data = np.load(file_path)
+        data = torch.tensor(data, dtype=torch.float32)
+
+        # Ensure shape is compatible with CNN (3 channels, 224x224)
+        if len(data.shape) == 2:
+            data = data.unsqueeze(0)
+
+        data = torch.nn.functional.interpolate(
+            data.unsqueeze(0), size=(224, 224), mode="bilinear"
+        ).squeeze(0)
+
+        data = data.repeat(3, 1, 1)
+
+        return data, label