LPIPS - Learned Perceptual Image Patch Similarity

This repository implements the LPIPS metric, a perceptual similarity metric for comparing images. The LPIPS metric is widely used in image generation and restoration tasks to evaluate the perceptual quality of images.

Features

• VGG-based Feature Extraction: Uses a pre-trained VGG16 model to extract multi-level features.
• Learned Linear Layers: Applies learned weights to compute perceptual similarity.
• Pre-trained Weights: Includes pre-trained weights for LPIPS computation.
• Utility Functions: Includes helper functions for downloading and verifying model checkpoints.

Installation

1. Clone the repository:

   git clone https://github.com/ProgramerSalar/LPIPS.git
   cd LPIPS

2. Install the required dependencies:

   pip install -r requirements.txt

Usage

Compute LPIPS Distance

You can compute the perceptual distance between two images using the LPIPS class:

import torch
from lpips import LPIPS

# Create sample images
img1 = torch.randn(1, 3, 256, 256)
img2 = torch.randn(1, 3, 256, 256)

# Compute perceptual distance
lpips = LPIPS().eval()
distance = lpips(img1, img2)
print(f"Perceptual distance: {distance.item():.4f}")

Load Pre-trained Model

To use the pre-trained LPIPS model:

from lpips import LPIPS

lpips_model = LPIPS.from_pretrained('vgg_lpips')
distance = lpips_model(img1, img2)
print(f"LPIPS distance: {distance.item():.3f}")

File Structure

Lpips/
├── .gitignore
├── assets/                # Folder for storing images and other assets (ignored by Git)
│   ├── download (1).jpeg
│   ├── download.jpeg
│   └── lpips_example.png  # Example image for the Readme (if added)
├── lpips.py               # Implements the LPIPS model and its components
├── utils.py               # Utility functions for downloading and verifying model checkpoints
├── vgg.pth                # Pre-trained weights for the LPIPS model
├── example.txt            # Example file for testing purposes
├── requirements.txt       # Dependencies for the project
└── Readme.md              # Project documentation

License

This project is licensed under the MIT License. See the LICENSE file for details.

References

• LPIPS: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric