From Structure to Detail: Hierarchical Distillation for Efficient Diffusion Model

The recent paper on Hierarchical Distillation (HD) presents a solution to the critical issue of inference latency in diffusion models, which has hindered their real-time application. Traditional methods, whether trajectory-based or distribution-based, have inherent trade-offs—trajectory methods preserve global structure but lose high-frequency details, while distribution methods achieve higher fidelity but face challenges like mode collapse. The HD framework synergistically integrates these approaches, using trajectory distillation to create a structural sketch that optimally initializes the distribution-based refinement stage. This innovative strategy not only enhances overall performance but also introduces an Adaptive Weighted Discriminator to improve adversarial training. The results are promising, demonstrating state-of-the-art performance across various tasks, particularly on ImageNet, indicating a significant leap forward in the efficiency and effectiveness of AI models.