Standard metric for image generation quality

WHAT THIS TESTS The interviewer wants to know whether you can judge a generative model objectively rather than by squinting at a few cherry-picked outputs. Image generation has no ground-truth target per sample, so the metric must compare distributions, not individual pixels.

A GOOD ANSWER COVERS Name Frechet Inception Distance as the de facto standard. Explain the mechanism: push a large set of real images and a large set of generated images through a pretrained Inception v3 network, take activations from a deep pooling layer, and model each set as a multivariate Gaussian described by a mean vector and covariance matrix. FID is the Frechet distance between those two Gaussians, combining how far apart the means are and how different the covariances are. Lower FID means generated images are statistically closer to real ones. A strong answer notes FID captures both fidelity, are samples realistic, and diversity, does the model cover the full data distribution rather than collapsing to a few modes.

COMMON WRONG ANSWERS Proposing pixel-level MSE or PSNR against real images, which assumes a paired target that does not exist. Suggesting only a grid of samples shown to humans, which is subjective and hides mode collapse. Confusing FID with Inception Score, which uses only generated images and no real reference.

LIKELY FOLLOW-UPS How many samples do you need, FID is biased upward on small sets, so use tens of thousands. Why Inception features, they encode semantic content. What are FID weaknesses, sensitivity to the feature extractor and to image preprocessing.

ONE CONCRETE EXAMPLE You train two GANs. Both produce sharp samples, but model A scores FID 12 and model B scores FID 45. Inspection reveals B keeps generating the same handful of faces. FID exposed B's mode collapse that a small sample grid hid.