ARIA Optimizes Conditional Diffusion Model Distillation

Distilling large conditional diffusion models into smaller, more efficient student models is a critical task, but it faces a significant challenge: maintaining alignment across a vast range of conditioning inputs. Unlike simpler recognition tasks, the noise predicted by diffusion models is highly dependent on the conditioning signal, meaning effective distillation requires exploring an extensive conditioning space. This becomes a bottleneck when paired image-condition data is scarce or when the pool of potential conditions, such as text prompts, is prohibitively large for exhaustive synthetic data generation. Existing methods attempt to broaden the conditioning space by switching conditions during training. However, this raises a new question: how should training effort be optimally distributed once a large conditioning corpus is available? This research introduces ARIA (Adaptive Region-Based Importance Allocation), a framework designed to address this by dynamically allocating training resources across coarse regions of the conditioning space. ARIA continuously estimates the discrepancy between the teacher and student models at a regional level. This allows it to prioritize training updates in areas where misalignment is most persistent, all while adhering to the original distillation objective. Empirical results show that ARIA consistently outperforms previous methods across various architectures and settings, with its most pronounced benefits observed in scenarios involving unseen or underrepresented conditioning regimes. The framework also includes a theoretical analysis supporting its mechanism for tracking evolving discrepancies during training.