New method enhances speculative sampling for diffusion models

VERA Brief

AI-generated. Grounded in the article and its cited sources.

Researchers have developed Rank-Aware Speculative Sampling (RASS) to accelerate diffusion model generation. RASS improves efficiency over existing tree-based methods by leveraging the ranking of draft candidates without additional target-model evaluations.

Key facts

• Rank-Aware Speculative Sampling (RASS) is a new method to accelerate diffusion model generation.
• RASS builds on tree-based speculative sampling techniques that generate and verify multiple draft candidates in parallel.
• The method orders candidates based on proposal-target mean displacement and samples a rank using optimized weights.
• Experiments show RASS improves upon previous methods like Diffusion Greedy Rejection Sampling (D-GRS).
• RASS achieved gains of approximately 20% on CIFAR-10 at comparable compute budgets.