← AI Terminology
FID - Fréchet Inception Distance
FID (Fréchet Inception Distance) compares distributions of real and generated images in Inception network feature space — lower is better.
It is the default automatic metric for generative image models.
It is the default automatic metric for generative image models.
Why It Matters in AI
Human eval does not scale for GANs/diffusion. FID correlates reasonably with quality/diversity tradeoffs and dominates papers and dashboards — with known limitations (Inception bias, not perfect perception).
Key Points
| Aspect | Description |
|---|---|
| Use | GANs, diffusion, latent generative models |
| Scale | Lower FID → closer distributions |
| Limits | Not fully aligned with humans; dataset dependent |
| Related | IS, precision/recall for generative models, CLIPScore |
| Practice | Report with same Inception/FID code versions |
| Computation | Fit Gaussians to real/fake features; Fréchet distance |
Simple Analogy
Comparing two clouds of perfume samples by chemistry lab stats — if the synthetic cloud matches the real cloud’s statistics, the fragrance line is close.
Common Usage Examples
pytorch-fidon sample folders- Track FID during diffusion training
- Report vs competing generators
- Complement with human preference studies
Summary
In short: FID measures how close generated image statistics are to real ones — the standard automatic score for image generators.