Metric & Benchmark
aka Fréchet Inception Distance
FID
Distributional distance between generated and real images, lower is better.
Definition
FID embeds real and generated images with an Inception-V3 network, fits Gaussians to each set's features and reports the Fréchet distance between them. It captures both diversity and fidelity and is the dominant metric for image-generation benchmarks.
Common use cases
- Image GAN/diffusion eval
- Model comparison