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

We present \textbf{AAD-1}, an \textbf{A}symmetric \textbf{A}dversarial \textbf{D}istillation framework for \textbf{O}ne-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.

Video generation models must synthesize realistic visual content while maintaining coherent motion over time. Recent distillation-based approaches can reduce the number of sampling steps required for generation, but they often remain difficult to train and may suffer from motion collapse, producing videos with limited or unnatural dynamics. This work introduces AAD-1, a framework for efficient one-step autoregressive image-to-video generation. AAD-1 uses an asymmetric adversarial design: the generator preserves the sequential structure required for autoregressive sampling, while the discriminator evaluates the full video sequence jointly. This enables the discriminator to assess global spatiotemporal consistency and detect failures such as frozen motion, inconsistent dynamics, and long-range drift. To improve training stability, AAD-1 adopts a phased strategy that first initializes the student model with distribution matching before applying adversarial distillation. This provides a stable starting point and makes subsequent adversarial training more reliable. Experiments on standard video generation benchmarks demonstrate that AAD-1 improves both visual quality and temporal consistency, achieving state-of-the-art performance in one-step autoregressive video generation. These results suggest that asymmetric adversarial distillation is an effective direction for building faster, more stable, and more practical video generation systems.