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

2026-06-02

AI Summary

Existing one-step autoregressive video generation methods often produce static outputs due to motion collapse and training instability. This work proposes an asymmetric adversarial distillation framework that mitigates these issues by breaking the architectural symmetry between generator and discriminator: it employs a causal autoregressive generator paired with a bidirectional spatiotemporal attention discriminator. Combined with a staged training strategy and a distribution-matching-based warm-up mechanism, the approach effectively alleviates motion collapse and enhances training stability. Evaluated on the VBench benchmark, the method achieves state-of-the-art performance, significantly improving both the dynamism and overall quality of generated videos.

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. Our key architectural insight is to break the symmetry between generator and discriminator. While 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 cause motion collapse in autoregressive generation. To stabilize training, we introduce a phased strategy that first uses distribution matching to bootstrap a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. Extensive experiments on VBench demonstrate that AAD-1 achieves state-of-the-art performance in one-step autoregressive video generation.

Problem

• motion collapse
• training instability
• autoregressive video generation
• adversarial distillation
• one-step generation

Innovation

• Asymmetric Adversarial Distillation
• One-step Autoregressive Video Generation
• Motion Collapse Mitigation
• Bidirectional Discriminator
• Phased Training Strategy