Towards One-step Causal Video Generation via Adversarial Self-Distillation

TL;DR

The paper tackles the challenge of efficient causal video generation with diffusion-based models by introducing Adversarial Self-Distillation (ASD) atop Distribution Matching Distillation (DMD), enabling high-quality synthesis with as few as $1$–$2$ denoising steps. It further introduces First-Frame Enhancement (FFE) to allocate more denoising effort to the first frame, mitigating error propagation while applying larger skip steps to later frames. A single, step-unified distilled model trained with ASD achieves strong 1- and 2-step performance and flexible deployment across various inference step settings, reducing the need for multiple distillations. Extensive experiments on VBench show superior quality and efficiency compared to baselines, with ablations confirming the contributions of ASD and FFE. The approach offers practical benefits for real-time, interactive video generation and scalable deployment across different resource budgets.

Abstract

Recent hybrid video generation models combine autoregressive temporal dynamics with diffusion-based spatial denoising, but their sequential, iterative nature leads to error accumulation and long inference times. In this work, we propose a distillation-based framework for efficient causal video generation that enables high-quality synthesis with extremely limited denoising steps. Our approach builds upon the Distribution Matching Distillation (DMD) framework and proposes a novel Adversarial Self-Distillation (ASD) strategy, which aligns the outputs of the student model's n-step denoising process with its (n+1)-step version at the distribution level. This design provides smoother supervision by bridging small intra-student gaps and more informative guidance by combining teacher knowledge with locally consistent student behavior, substantially improving training stability and generation quality in extremely few-step scenarios (e.g., 1-2 steps). In addition, we present a First-Frame Enhancement (FFE) strategy, which allocates more denoising steps to the initial frames to mitigate error propagation while applying larger skipping steps to later frames. Extensive experiments on VBench demonstrate that our method surpasses state-of-the-art approaches in both one-step and two-step video generation. Notably, our framework produces a single distilled model that flexibly supports multiple inference-step settings, eliminating the need for repeated re-distillation and enabling efficient, high-quality video synthesis.

Figures (9)

• Figure 1: The different alignment strategies of DMD and ASD during distillation. $p_{\text{n-gen}}$ represents the generation distribution of the $n$-step distillation model. The points plotted in the figure indicate the multiple modes (or peaks) of the distribution, highlighting its multimodal nature.
• Figure 2: Qualitative results of Self Forcing huang2025self and ours under 4-step, 2-step and 1-step generation. Our method consistently maintains high-quality generation across 4-step, 2-step, and 1-step inference.
• Figure 3: Pipeline of our proposed adversarial self-distillation process. We employ a discriminator $D_n$ to align the randomly noised $n$-step video with the $(n\!+\!1)$-step one through calculating the ASD loss in \ref{['eq:consistency_loss']}. The generator $G$ is optimized using a combined objective function that includes the standard DMD loss and the ASD loss. Note that distillation is performed in the latent space, while the pixel domain is primarily used for visual analysis and display.
• Figure 4: Cosine similarity matrices of different frames in causal diffusion video generation. Each matrix shows the similarity of the predicted $\hat{x}_0$ between different denoise steps (from 1 to 4).
• Figure 5: Qualitative comparisons. We visualize videos generated by Ours against those by Wan2.1 wang2025wan, SkyReels-V2 chen2025skyreels, CausVid yin2025causvid and Self Forcing huang2025self at 4-step generation. All models share the same architecture with 1.3B parameters.