The problem — Generative models for synthetic microscopy data — GANs for virtual staining, diffusion models for Cell Painting augmentation, domain-adaptation networks for H&E normalization — need a scalar readout that goes beyond per-image pixel similarity. SSIM and PSNR require a pixel-paired reference; they say nothing about whether a generator reproduces the diversity and distributional shape of real biology across thousands of cells. Teams reach for FID to fill that gap. Before trusting the number, it is worth understanding what it measures and, more precisely, what it does not.
What it is / how it works — FID computes the Fréchet distance between two multivariate Gaussians: one fit to Inception-v3 pool3 features extracted from a set of real images, one fit to features from generated images. Lower FID means the two distributions are closer in embedding space. Because it operates on feature distributions rather than pixel pairs, it captures global realism — mode coverage, diversity, and texture coherence — in a single number. That makes it the standard evaluation metric for generative model development, and increasingly the entry-level check when validating synthetic training data or stain-normalization pipelines.
Where it breaks — FID carries three structural limitations that matter specifically in microscopy. First, the Inception-v3 backbone was trained on ImageNet natural images, not biological specimens: the features it responds to (texture, object boundaries, color) do not map cleanly onto the structures that matter for quantitative microscopy — organelle morphology, intensity periodicity, nuclear texture. A generator can score well on FID while corrupting the feature space a downstream segmenter or profiling model actually reads. Second, FID can be gamed: a generator that reproduces each feature's marginal distribution while destroying inter-feature dependencies will achieve a low score yet produce images that are biologically incoherent. Third, FID is biased and unstable at small sample sizes — reliable estimation typically requires several thousand images, which is not always available for rare phenotypes or niche assays.
The Fovea position: FID belongs at the readout-distribution level of the quality hierarchy described in Image Quality Is Not One Number, not at the image-space level where SSIM and PSNR operate. It is a necessary but insufficient check. For synthetic microscopy data to be trustworthy in a production pipeline, complement FID with biology-conditioned metrics: profile reproducibility (percent-replicating, mAP) on the synthetic cohort, segmentation accuracy of a fixed model on generated versus real images, and downstream classification or readout concordance. A generator that passes FID but degrades profile reproducibility has moved the problem, not solved it.