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

AAD-1, from SJTU, Ant Group, Tsinghua University, Zhejiang University, and Anyverse Dynamics, tackles one-step autoregressive image-to-video generation by breaking the symmetry between generator and discriminator: the causal generator preserves streaming inference, while a bidirectional discriminator with a video-level holistic logit can see the full spatiotemporal rollout and detect motion collapse and long-range drift that causal critics miss. The method combines a three-stage training recipe—ODE initialization, a self-rollout distribution-matching warm-up, and asymmetric adversarial refinement—on top of the public Wan 2.1 T2V 14B backbone, using sink tokens, a sliding window of 9, and a single sampling step per chunk. On VBench-I2V, the final one-step model reaches 94.34 subject consistency, 95.08 background consistency, 98.22 motion smoothness, and 98.65 I2V subject faithfulness, outperforming prior autoregressive baselines such as CausVid and Self Forcing while matching or exceeding the bidirectional Wan 2.1 reference on several consistency metrics. The ablations show why the design matters: without DMD warm-up, aesthetic quality drops to 53.63 and imaging quality to 62.81, while a causal frame-wise discriminator collapses to static video with a Dynamic Degree of 1.08; the best regularization is λ = 20, because λ = 0 causes training collapse and λ = 50 introduces grid artifacts. The paper also reports that full training takes about 3.5 days on 64 NVIDIA H20 GPUs, and one-step inference on a 14B model runs at 1.134 s and 14.33 FPS on a single H100 GPU.