Information-Theoretic CFG Optimizes Diffusion Model Guidance Schedules

A new information-theoretic framework has been proposed to optimize Classifier-Free Guidance (CFG) schedules in diffusion models. CFG is crucial for conditional generation, enhancing consistency but often at the cost of diversity and distributional coverage. The challenge lies in dynamically controlling this trade-off across the generation trajectory. The new method addresses this by using a clean endpoint reference to define the desired consistency-coverage balance. It then optimizes the actual distribution produced by the guided sampler to align with this reference. By deriving trajectory-level formulas, the objective can be estimated from samples and score evaluations without explicit density estimation. This approach has shown competitive or improved trade-offs on ImageNet-512 and COCO with SD-XL, demonstrating its ability to selectively allocate guidance across different noise levels.