Evaluation Metrics for Conditional Image Generation

We present two new metrics for evaluating generative models in the class-conditional image generation setting. These are obtained by generalizing most popular unconditional metrics: Inception Score (IS) and Fre'chet Distance (FID). A theoretical analysis shows motivation behind each proposed metric links novel to their counterparts. The link takes form of a product case IS or an upper bound FID case. provide extensive empirical evaluation, comparing variants other metrics, utilize them analyze existing models, thus providing additional insights about performance, from unlearned classes mode collapse.