Visualizing Distribution Coverage in Generative Diffusion Models
Published in arXiv preprint, 2026
Single-sample evaluations can hide whether a generator explores the full range of valid outputs. We use pass@k to measure success as the sampling budget grows, finding that distribution-matching objectives improve early-hit rates while reducing large-budget coverage. Consistency and trajectory-based distillation better preserve the teacher’s coverage, and the tradeoff also appears in causal video generation.
Recommended citation: Wang, Y., Wu, X., Fu, T.-J., Chen, C., Chen, L.-C., Gan, Z., and Wei, C. (2026). "Visualizing Distribution Coverage in Generative Diffusion Models." arXiv:2609.38853.
Download Paper