fid

**FID** is the **Frechet Inception Distance metric that compares feature distributions of generated images and real images to estimate realism gap** - it is one of the most widely used generative-image evaluation metrics. **What Is FID?** - **Definition**: Distribution-distance score computed between Gaussian approximations of deep feature embeddings. - **Feature Source**: Typically uses activations from a pretrained Inception network layer. - **Interpretation**: Lower FID indicates generated image distribution is closer to real data distribution. - **Usage Scope**: Common in GAN and diffusion-model benchmarking across datasets. **Why FID Matters** - **Standard Benchmark**: Provides shared quantitative baseline for generative model comparison. - **Distribution Focus**: Captures realism and diversity jointly at dataset level. - **Regression Tracking**: Useful for monitoring generation quality drift across training runs. - **Research Communication**: Widely reported metric supports cross-paper comparability. - **Caveat Awareness**: Sensitive to sample count, preprocessing, and domain mismatch. **How It Is Used in Practice** - **Protocol Consistency**: Use fixed preprocessing and sufficient sample size for stable comparisons. - **Complementary Metrics**: Pair FID with human studies and prompt-alignment scores for fuller evaluation. - **Reproducibility Controls**: Document seeds, dataset splits, and evaluation code versions. FID is **a central distribution-based metric in generative vision evaluation** - FID is most useful when computed with strict, reproducible evaluation protocol.