Advanced
Open
Pro
FID vs. Inception Score: What Would Move Each
Part of the ML System Design Interview path →
Part of the Generative Vision & Image AI System Design path →
Two proposed changes to your image generation system are under debate:
(a) Add a diversity-encouraging sampling tweak that makes the model occasionally generate unusual, less-common compositions it rarely produced before, at a slight cost to average per-image sharpness. (b) Fix a bug where the tokenizer's decoder was slightly blurring fine texture across the board, with no effect on which types of scenes get generated.
- Explain what Inception Score and FID each actually measure.
- Predict the likely effect of change (a) and change (b) on each metric, and justify your prediction from what the metric measures.
- Why might a team want both metrics tracked rather than picking one?
Share this question