Hi Rebecca,

Thank you for your question.

What is FID?

Fréchet Inception Distance (FID) is a widely adopted metric for evaluating the performance of generative models, particularly in image generation tasks.

It works by:

• Extracting feature representations of both real and generated images using a pre-trained Inception network.
• Treating these features as samples from multivariate Gaussian distributions.
• Calculating the Fréchet distance between the two distributions.

A lower FID score indicates higher similarity between the real and generated image distributions.

A score of 0 means the two sets are identical in the feature space.

Why is FID Better than Other Metrics?

FID offers several key advantages over traditional evaluation metrics like Inception Score (IS), SSIM, or PSNR:

• Considers Both Real and Generated Data

Unlike IS, which only evaluates generated images, FID compares distributions from both real and generated samples, leading to a more comprehensive evaluation.

• Sensitive to Mode Collapse

FID increases when a generative model lacks diversity, even if image quality is high—making it effective for identifying mode collapse issues.

• Correlates with Human Perception

Research shows that FID scores align well with human judgments of image realism, making the metric more meaningful in real-world applications.

• Efficient and Stable

Once features are extracted, FID only requires the computation of means and covariances. No additional models or complex scoring logic are needed.

• Supports Fair Cross-Model Comparison

The use of a fixed feature extractor ensures consistency across different datasets and model architectures, enabling fair benchmarking.

In Summary

FID is a reliable, efficient, and interpretable metric that captures both the realism and diversity of generated images.

It offers more meaningful insights than traditional metrics, making it a preferred choice in the evaluation of generative models.