GPD: Guided Progressive Distillation for Fast and High-Quality Video Generation

TL;DR

The paper tackles the high computational cost of diffusion-based video generation by proposing Guided Progressive Distillation (GPD), which uses online teacher refinement and progressive step-size training to enable accurate one-step predictions with larger horizons. It combines a three-model interaction to generate online training targets and a frequency-domain high-frequency loss to preserve details and temporal dynamics. On Wan2.1, GPD reduces sampling steps from 48 to 6 and achieves an 8× speedup while attaining state-of-the-art VBench performance (84.04%), outperforming PeRFlow and CausVid baselines. The approach is data-efficient, requiring only textual prompts, and demonstrates robust transfer across resolutions, offering a practical route to fast, high-quality video synthesis.

Abstract

Diffusion models have achieved remarkable success in video generation; however, the high computational cost of the denoising process remains a major bottleneck. Existing approaches have shown promise in reducing the number of diffusion steps, but they often suffer from significant quality degradation when applied to video generation. We propose Guided Progressive Distillation (GPD), a framework that accelerates the diffusion process for fast and high-quality video generation. GPD introduces a novel training strategy in which a teacher model progressively guides a student model to operate with larger step sizes. The framework consists of two key components: (1) an online-generated training target that reduces optimization difficulty while improving computational efficiency, and (2) frequency-domain constraints in the latent space that promote the preservation of fine-grained details and temporal dynamics. Applied to the Wan2.1 model, GPD reduces the number of sampling steps from 48 to 6 while maintaining competitive visual quality on VBench. Compared with existing distillation methods, GPD demonstrates clear advantages in both pipeline simplicity and quality preservation.

Figures (5)

• Figure 1: Motivation of Guided Progressive Distillation (GPD). Traditional flow-straightening forces a student to approximate the teacher’s highly curved multi-step sampling trajectory in a single step, often causing large target mismatch and low-quality predictions. We propose teacher refinement, where the teacher further refines the student’s one-step intermediate output to form an adaptive GPD target that is both high-fidelity and easier to learn than a constant target. As shown by the vector alignment analysis and cosine similarity across timesteps, the refined target’s update direction aligns better with the student’s current inference flow.
• Figure 2: Overview of Guided Progressive Distillation. Progressive distillation: training is split into stages that gradually extend the student’s step size. Teacher-refined target: the teacher refines an intermediate latent predicted by the frozen prior student to form the supervision target. High-frequency loss: a 3D FFT high-pass penalty enforces matching high-frequency details between the student output and the teacher-refined target.
• Figure 3: Timestep Hierarchy in Frequency Information Reconstruction. Initially, high-frequency components appear as unstructured noise, but as $t$ decreases, they reveal meaningful structures such as edges and motion details.
• Figure 4: Qualitative results on text-to-video. Qualitative comparison showing our method achieves 8× faster inference than Wan2.1-1.3B while maintaining superior visual quality. Compared to PeRFlow, our approach eliminates noticeable distortions in subjects and backgrounds. Compared to CausVid, our method maintains stronger semantic coherence and preserves fine-grained elements.
• Figure 5: Online refinement training improves distillation quality. Compared with conventional training using a constant target, online refinement post-training yields consistent gains in the overall score and across multiple sub-metrics, indicating improved visual fidelity and stronger temporal consistency.