Fréchet Inception Distance (FID)

FID measures the Wasserstein-2 distance (Fréchet distance) between two multivariate Gaussian distributions fitted to the feature embeddings of real and generated images. It assumes the 2048-dimensional activations from the Inception v3 pool3 layer are normally distributed. A lower FID indicates that the synthetic images are statistically closer to the real distribution in terms of both mean and covariance.

Unlike the Inception Score (IS), FID is highly sensitive to mode collapse—a failure mode where a generator produces only a limited variety of samples. Because FID evaluates the covariance structure of the generated distribution, a model that produces a single high-quality image repeatedly will exhibit a poor (high) FID, as its feature variance will be near zero compared to the diverse real distribution.

FID relies on the Inception v3 network pre-trained on ImageNet as a fixed feature extractor. This introduces a known bias: the metric assumes that features relevant to ImageNet classification are also relevant for perceptual image quality. For specialized domains like industrial defect detection or medical imaging, this assumption may not hold, leading researchers to compute domain-specific variants using custom feature extractors.

FID is a biased estimator; its calculated value depends heavily on the number of samples used. The original authors recommend a minimum of 10,000 samples for a reliable estimate, with 50,000 being standard. Using fewer samples produces an upward bias, making models appear worse than they are. Always report the sample size alongside the FID score to ensure reproducibility and fair comparison.

FID correlates better with human judgment of visual fidelity than pixel-space metrics like Mean Squared Error (MSE) or Peak Signal-to-Noise Ratio (PSNR). However, it is not a perfect proxy for human perception. A low FID does not guarantee that individual images are free of artifacts; it only confirms that the aggregate statistical properties of the generated set match the real set in the Inception feature space.

In manufacturing contexts, FID is used to validate synthetic defect generation pipelines. A low FID between a synthetic defect library and real defect images suggests the artificial data captures the essential texture, lighting, and geometry of actual production anomalies. This validation step is critical before using synthetic data to train computer vision quality inspection models, ensuring the domain gap is quantified.

The predecessor to FID, also using the Inception network. IS evaluates synthetic images based on two criteria: low entropy conditional label distributions (each image clearly belongs to one class) and high entropy marginal label distribution (diverse outputs across classes). Unlike FID, IS does not compare against real data statistics, making it blind to mode collapse where a generator produces only one highly confident but repetitive output. FID is generally preferred for industrial defect generation because it directly measures distributional similarity to real defect samples.

The statistical divergence between the feature distributions of synthetic training data and real-world operational data that degrades model performance upon deployment. FID quantifies this gap directly: a high FID between synthetic and real images indicates a large domain gap. In sim-to-real transfer for manufacturing inspection, minimizing FID through techniques like domain randomization and photorealistic rendering is critical. A low FID score correlates with better downstream task performance when models trained on synthetic data are deployed on real production lines.

A measure of how closely a synthetic dataset statistically mirrors the properties and distributions of the real-world data it is intended to replace or augment. FID is the de facto standard for quantifying fidelity in image generation. Key aspects include:

• Visual realism: Per-pixel and textural similarity
• Distributional alignment: Matching feature statistics via FID
• Diversity coverage: Avoiding mode collapse In industrial defect generation, high fidelity ensures that quality inspection models trained on synthetic anomalies generalize effectively to real production environments.

FID relies on the Inception-v3 convolutional neural network pre-trained on ImageNet as a fixed feature extractor. Images are passed through the network up to the final average pooling layer, producing a 2048-dimensional feature vector for each image. These embeddings capture high-level semantic and textural properties. The FID score is then computed as the Fréchet distance between multivariate Gaussians fitted to the real and synthetic feature sets. This reliance on a natural-image-trained backbone is a known limitation when evaluating specialized industrial imagery like thermal or X-ray scans.