inception score

**Inception score** is the **generative-image metric that measures confidence and diversity using class-probability outputs from an Inception classifier** - it was an early benchmark for GAN quality evaluation. **What Is Inception score?** - **Definition**: Score based on KL divergence between conditional class distribution and marginal class distribution. - **Intuition**: High confidence per image and diverse classes across images produce higher score. - **Computation Basis**: Relies on pretrained classifier predictions rather than direct human judgments. - **Historical Role**: Widely used before broader adoption of FID and newer perceptual metrics. **Why Inception score Matters** - **Diversity Signal**: Rewards output sets that cover multiple semantic categories. - **Quality Proxy**: Penalizes blurry or ambiguous images that produce uncertain classifier outputs. - **Benchmark Legacy**: Still appears in literature and historical model comparisons. - **Limitations Insight**: Does not compare against real data distribution directly. - **Evaluation Context**: Useful only when interpreted with known constraints and complementary metrics. **How It Is Used in Practice** - **Protocol Clarity**: Report exact classifier setup and preprocessing for comparability. - **Metric Pairing**: Combine with FID and human preference studies to offset blind spots. - **Domain Check**: Avoid over-reliance when generated data differs from classifier training domain. Inception score is **an important historical metric for generative-image benchmarking** - Inception score should be used with caution and complementary evaluation methods.