What determines the quality of synthetic data?

Mostly the input dataset's size and characteristics — bigger, well-behaved datasets with few outliers and simpler relationships yield higher-quality synthetic data.

A generative model can only reproduce patterns it could reliably learn, so synthetic-data quality is really a question of how much trustworthy signal the original dataset offers. The dominant factor is sheer size: a small set (<1,000 rows) has too little variability for stable statistical patterns, a medium one (1,000–10,000) supports basic modeling, and a large one (>10,000) is ideal because rich variability lets the model capture complex relationships — more individuals simply means more robust patterns to imitate.

Beyond raw size, the shape of the data matters. Well-behaved distributions (roughly bell-shaped, single-mode) are far easier to model than heavily skewed or multi-modal ones, where a simple model will smooth away real structure. Outliers are a particular trap — a few extreme values distort the summary statistics the model learns, so it generates unrealistic records. And the more complex the interactions between variables, the more sophisticated a generator you need to reproduce them faithfully rather than inventing impossible combinations.

Tip: Synthetic data and differential privacy share the same friend — big datasets. Both struggle when there's too little signal to learn from without either over-fitting individuals or drowning them in noise.