Evaluating Generative Models via One-Dimensional Code Distributions
The episode discusses a paper evaluating generative models using discrete visual tokens instead of continuous features. Hosts explain how Codebook Histogram Distance (CHD) and Code Mixture Model Score (CMMS) measure image quality by analyzing token frequencies and spatial co-occurrences. The discussion concludes that this token-based approach offers a more faithful, training-free, and sample-efficient way to assess model performance compared to older continuous feature metrics.