Optimizing Few-Step Generation with Adaptive Matching Distillation

Lichen Bai, zikai zhou, Shitong Shao, Wenliang Zhong, Shuo Yang, Shuo Chen, Bojun Cheng, Zeke Xie

TL;DR: We introduce Adaptive Matching Distillation (AMD), a reward-aware framework that dynamically adapts the teacher signals to escape "Forbidden Zones" and stabilize few-step diffusion generation.

Abstract: Distribution Matching Distillation (DMD) is a powerful acceleration paradigm, yet its stability is often compromised in **Forbidden Zones**—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.

Lay Summary:

Making AI image and video generators faster often involves a technique that learns from two guides: a reliable one (the “real teacher”) and a less reliable one (the “fake teacher”). However, this process can become unstable in certain tricky situations—called Forbidden Zones—where the real teacher gives confusing advice and the fake teacher fails to push the model back on track.

In this work, we first show that previous methods actually tried to avoid these zones without directly addressing them. Then we introduce a new, smarter approach called Adaptive Matching Distillation (AMD). AMD acts like a self-correcting system: it constantly checks for signs of trouble, escapes from Forbidden Zones when needed, and actively builds protective barriers to prevent the model from collapsing into bad behaviors.

We tested AMD on popular image and video generation tasks (like SDXL and Wan2.1) and on standard quality benchmarks. The results show clear improvements in both image quality and training stability. For example, AMD boosted a key quality score from 30.64 to 31.25 on SDXL, outperforming other leading methods. This demonstrates that directly fixing the optimization path inside Forbidden Zones is the key to pushing fast generative models to new heights.

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: Diffusion Model, Step Distillation, Video Generation

Submission Number: 10939