AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Image-to-Video Generation

Forum topic · 小凯 · 2026-06-04

Summary

AAD-1 is an asymmetric adversarial distillation framework for one-step autoregressive image-to-video generation, presented in arXiv paper 2606.03972. State-of-the-art adversarial distillation methods for video generation suffer from motion collapse and training instability, often producing static videos. AAD-1 tackles these issues through two key designs. Architecturally, it breaks the symmetry between generator and discriminator: the generator stays causal to preserve autoregressive sampling capability, while the discriminator attends bidirectionally over the full spatiotemporal context and outputs a single holistic realism score for the entire video sequence. This asymmetry lets the discriminator detect global temporal failures and long-range drift that cause motion collapse in autoregressive generation. For training stability, AAD-1 introduces a staged strategy that first uses distribution matching to build a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. The paper is authored by Haobo Li, Yanhong Zeng, Yunhong Lu, and colleagues.

Overview

Field: Computer Vision (CV) Authors: Haobo Li, Yanhong Zeng, Yunhong Lu, Jiapeng Zhu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Yujun Shen, Zhipeng Zhang arXiv: 2606.03972

Abstract

We present AAD-1, an Asymmetric Adversarial Distillation framework for One-step autoregressive image-to-video generation. State-of-the-art methods adopt adversarial distillation but suffer from motion collapse and training instability, resulting in static videos. AAD-1 addresses these challenges through two key designs in architecture and training strategy.

Asymmetric Architecture

The key architectural insight is to break the symmetry between generator and discriminator:

The generator remains causal to preserve autoregressive sampling capability.
The discriminator attends bidirectionally over the full spatiotemporal context and produces a single holistic realism score for the entire video sequence.

This asymmetric design enables the discriminator to effectively detect global temporal failures and long-range drift that lead to motion collapse in autoregressive generation.

Staged Training Strategy

To stabilize training, AAD-1 introduces a staged strategy:

1. First use distribution matching to guide a stable one-step generator. 2. This warm-up phase brings the student distribution closer to the teacher before adversarial distillation begins.