Optimizing Few-Step Generation with Adaptive Matching Distillation
Abstract
Adaptive Matching Distillation introduces a self-correcting mechanism to improve generative model training by detecting and escaping unstable regions in the optimization landscape.
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.
Community
This paper proposes Adaptive Matching Distillation (AMD), a unified optimization framework that significantly stabilizes and improves few-step diffusion distillation (e.g., SDXL, Wan2.1).
💡 Key Insight: The "Forbidden Zone"
The authors identify a critical failure mode in DMD called the "Forbidden Zone"—regions where the real teacher provides unreliable guidance (hallucinated gradients) while the fake teacher exerts insufficient repulsive force, leading to training collapse.
✨ What's New in AMD?
Instead of static distillation, AMD acts as a self-correcting system:
- Reward-Aware Diagnosis: Uses reward models as a proxy to detect samples trapped in Forbidden Zones.
- Dynamic Score Adaptation: Decomposes gradients into Distribution Matching and Conditional Alignment terms. It dynamically prioritizes repulsive forces to "kick" the student out of failure modes when samples are poor, and focuses on refinement when samples are good.
- Repulsive Landscape Sharpening: Forces the fake teacher to learn a steeper energy landscape around low-quality samples, amplifying the escape signal.
📈 Results
- Image Generation: AMD outperforms DMD2, PCM, and DMDR on SDXL (HPSv2 score improved from 30.64 → 31.25).
- Video Generation: Validated on Wan2.1, achieving superior motion smoothness and visual fidelity compared to standard baselines.
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper