Evaluation of Image Generation

FID, LPIPS, SSIM, KID…

1. Frechet Inception Distance (FID):

• Purpose: Measures the similarity between the PAIR of real and generated images regarding feature distributions.
• How it Works:
  Extract features from images using a pre-trained Inception network.
  Calculate the mean and covariance of these features for both real and generated images.
  Compute the Wasserstein-2 distance between these distributions.
  Lower FID indicates higher similarity (better quality).
• Strengths:
  Captures both fidelity and diversity.
  Sensitive to mode collapse.
• Weaknesses:
  Computationally expensive.
  Assumes Gaussian feature distribution.

!pip install torchmetrics[image]
from torchmetrics.image.fid import FrechetInceptionDistance

fid = FrechetInceptionDistance(normalize=True)
fid.update(real_images, real=True)
fid.update(fake_images, real=False)

2. Learned Perceptual Image Patch Similarity (LPIPS):

• Purpose: Measures perceptual similarity between image pairs.
• How it Works:
  Compare images at the feature level using a pre-trained deep network (e.g., VGG).
  Compute differences in feature maps and aggregate them spatially.
  Lower LPIPS indicates higher similarity (better perceptual quality).
• Strengths:
  Correlates well with human perception.
  Can be used to evaluate paired or unpaired datasets.