Search papers, labs, and topics across Lattice.
This paper addresses the instability of Distribution Matching Distillation (DMD) in generative models, particularly within "Forbidden Zones" where teacher guidance is unreliable. They introduce Adaptive Matching Distillation (AMD), a self-correcting mechanism that uses reward proxies to detect and escape these zones by prioritizing corrective gradients and sharpening energy barriers. AMD demonstrates improved sample fidelity and training robustness on image and video generation tasks, achieving a higher HPSv2 score on SDXL compared to existing methods.
By explicitly detecting and escaping "Forbidden Zones" during training, AMD unlocks significant gains in sample fidelity and training robustness for few-step generative models like SDXL.
Distribution Matching Distillation (DMD) is a powerful acceleration paradigm, yet its stability is often compromised in Forbidden Zone, regions where the real teacher provides unreliable guidance while the fake teacher exerts insufficient repulsive force. In this work, we propose a unified optimization framework that reinterprets prior art as implicit strategies to avoid these corrupted regions. Based on this insight, we introduce Adaptive Matching Distillation (AMD), a self-correcting mechanism that utilizes reward proxies to explicitly detect and escape Forbidden Zones. AMD dynamically prioritizes corrective gradients via structural signal decomposition and introduces Repulsive Landscape Sharpening to enforce steep energy barriers against failure mode collapse. Extensive experiments across image and video generation tasks (e.g., SDXL, Wan2.1) and rigorous benchmarks (e.g., VBench, GenEval) demonstrate that AMD significantly enhances sample fidelity and training robustness. For instance, AMD improves the HPSv2 score on SDXL from 30.64 to 31.25, outperforming state-of-the-art baselines. These findings validate that explicitly rectifying optimization trajectories within Forbidden Zones is essential for pushing the performance ceiling of few-step generative models.
Microsoft Research