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

1 Introduction

Diffusion models [Ho et al., 2020; Song et al., 2021] have emerged as a powerful framework for video generation [Blattmann et al., 2023; Zhou et al., 2023; Wan et al., 2025; Wang et al., 2024; Kong et al., 2025; Ma et al., 2025], achieving state-of-the-art results in temporal consistency and visual quality. However, video diffusion models require iterative denoising of Gaussian noise to generate final videos, making them computationally intensive and time-consuming.

To accelerate the diffusion process, traditional flowstraightening methods [Liu et al., 2022; Liu et al., 2024; Yan et al., 2024; Ke et al., 2025] force the student to cover the highly curved multi-step trajectory of the teacher in one step. This objective is inherently challenging and often leads to poor alignment, where the student's one-step prediction deviates significantly from the teacher's target, resulting in low-quality outputs as shown in Figure 1. We find that incorporating an additional teacher refinement step, based on the student's one-step intermediate prediction, can be beneficial. Specifically, it offers two key advantages. First, the refinement step produces high-quality supervision targets. By refining the student's initial prediction, the teacher provides high-fidelity signals that guide the student to produce videos with superior visual details and temporal consistency. Second, compared to the constant targets used in conventional methods, the teacher-refined target is adaptive and therefore easier to learn. As evidenced by the higher cosine similarity in Figure 1, the update direction of the refined target aligns more consistently with the student's current inference flow. In summary, this teacher refinement strategy effectively bridges the gap between generation quality and training difficulty, ensuring that the student learns from a target that is both highfidelity and structurally aligned with its inference path.

∗Corresponding Author.

Based on this insight, we further extend the concept of progressive optimization [Lin et al., 2024; Salimans and Ho, 2022] and incorporate quality-preserving techniques, proposing Guided Progressive Distillation (GPD), an acceleration framework that achieves efficient, high-quality video generation. First, we introduce online refinement training, where the teacher model dynamically refines the student's intermediate predictions during training, creating adaptive targets that follow straighter trajectories compared to teacher synthesis. Second, we design a new progressive distillation strategy adapted to online refinement training, enabling fast and stable convergence. Compared to other distillation methods, it substantially reduces training cost. Third, we incorporate highfrequency preservation through a frequency-domain loss that explicitly maintains fine-grained details and motion dynamics that are often smoothed out during trajectory straightening.

Extensive experiments demonstrate that GPD achieves state-of-the-art performance, generating high-quality videos in just 6 steps compared to the original 48 steps-an 8× speedup. On the VBench benchmark [Huang et al., 2023], our method achieves a total score of 84.04%, outperforming both traditional full-step models and existing acceleration methods including CausVid [Yin et al., 2025] (83.65%) and AccVideo [Zhang et al., 2025] (83.28%), while maintaining superior text alignment and visual fidelity across diverse video generation tasks.

2 RelatedWork

Video diffusion distillation methods can be broadly classified into two categories according to training objective: Distribution Matching methods and Trajectory Straightening methods.

Distribution Matching methods aim to align the output distribution of the student model with that of the teacher model. These methods leverage the score function to minimize the divergence of the distributions between teacher and student output [Luo et al., 2024; Yin et al., 2024b; Chadebec et al., 2024; Sauer et al., 2023; Yin et al., 2024a; Lin et al., 2024]. Recent advances include [Yin et al., 2025], which utilizes a teacher-synthesized dataset to train a onestep generator and extends Distribution Matching Distillation (DMD) as a post-training stage to enhance generation quality. However, these approaches rely on real video datasets for distribution matching, and the generated captions often lack fine-grained detail, which can weaken text-video alignment and reduce controllability.

Trajectory Straightening methods aim to straighten the probability flow trajectory by forcing the student model to predict the same target as the teacher's multi-step prediction. Consistency Model [Song et al., 2023] leverages a novel training paradigm that enables the model to predict the initial image at any timestep of the diffusion process, ultimately enabling training of a one-step generator. [Wang et al., 2023; Luo et al., 2023b] extended Consistency Model to latent space, facilitating high-resolution image and video synthesis with minimal inference steps.

Flow-based methods [Lipman et al., 2023] accelerate diffusion models by straightening the velocity field predicted by diffusion model. [Liu et al., 2022; Liu et al., 2024] introduce Rectified Flow to enhance both efficiency and transferability in data generation tasks. The concept was further expanded by [Esser et al., 2024; Yan et al., 2024; Ke et al., 2025], enabling efficient high-resolution image synthesis through transformer architectures. While these methods are easy to train using a simple regression loss, they often result in blurry outputs during few-step generation, especially when applied to video synthesis.

3 Background

3.1 FlowMatching for DiffusionModels

Diffusion models generate data by learning to reverse a gradual noising process [Ho et al., 2020; Song et al., 2021]. Under the flow-matching paradigm [Lipman et al., 2023], this process is formulated as learning a time-dependent velocity field vθ(zt, t) that transports samples along a probability flow ODE:

where zt represents the noisy data at time t, with z0 being clean data and zT being pure noise.

In practice, solving the ODE requires discretization. Let Sθ(zti , ti, tj) denote the ODE solver that integrates the velocity field vθ from time ti to tj:

For a teacher model with parameters ϕ, the generation process consists of N sequential denoising steps:

High-quality generation typically requires many evaluations (e.g., 50-100 steps) of the velocity field, which becomes computationally prohibitive, especially for video generation.

3.2 Distillation for Acceleration

Distillation methods aim to train a student model with parameters θ that can generate high-quality samples in K ≪ N steps. The rectified flow distillation approach, which we refer to as offline distillation, attempts to directly match the student's k-step prediction with pre-computed teacher trajectories [Zhang et al., 2025]:

where S (k) ϕ (zti ) denotes the teacher's k-step trajectory.

4 Method

4.1 Guided Progressive Distillation

Traditional distillation approaches employ the loss Lstandard with pre-computed teacher trajectories. The student model is required to match fixed targets sampled by the teacher along a highly curved trajectory in the latent space. This formulation suffers from two major shortcomings: (1) when k is large, the teacher trajectory becomes highly complex due to the accumulation of nonlinear transformations. Such high-curvature trajectories are difficult for the student to approximate within a single prediction step; (2) during training, inputs are generated solely by the teacher. At inference, errors from intermediate student predictions may accumulate, causing a distribution shift that remains unaddressed.

Progressive distillation with linear increment. To address the challenge of high trajectory curvature, we decompose the optimization process into K progressive stages, linearly increasing the prediction horizon of the student model. Assuming both the teacher and student share an initial unit step size, at each stage k ∈ 2, ...,K, we train the student to approximate a target with a step size of k. By gradually increasing the step size from 1 to K across multiple stages, this approach ensures a smooth transition in trajectory curvature, making the learning landscape tractable.

Instead of requiring the student to immediately bridge a complex K-step gap involving highly nonlinear trajectories, the progressive strategy enables the model to incrementally learn longer predictions. Each stage builds upon the stable foundation established by the previous one, allowing the student to adapt to increasing complexity in a structured and efficient manner.

Online training target with teacher guidance. To address the distribution shift between training and inference, we propose a dynamic online target generation mechanism. Instead of relying on static, precomputed teacher trajectories, our method constructs training targets dynamically via a three-model interaction process. This design allows the training objective to adapt continuously to the evolving capabilities of the student model.

Specifically, we use a frozen copy of the student model, vθk-1 , from the previous training stage to estimate the velocity field over a k - 1-step interval from the initial latent state zti . An ODE solver integrates this velocity to yield an intermediate latent zti-k+1 .

This intermediate state is then passed to the teacher model vϕ, which refines the trajectory by computing the velocity for the final step.z∗

The resulting refined latent state z∗ ti-k is used to compute a global average velocity over the full k-step interval:

This dynamically derived velocity serves as the optimization target for the current student model vθ. Such online target generation enhances robustness by enabling the teacher to correct the student's trajectory progressively.

The training objective for trajectory straightening at stage k is defined as:

This progressive, teacher-refined training pipeline combines the benefits of gradual time-step extension with online, adaptive supervision. As a result, the student learns to generalize across a wider temporal horizon, with improved resilience to distributional discrepancies between training and inference.

4.2 High-Frequency Preservation

High-frequency details are crucial for video visual quality, detail richness, and dynamic degree. We introduce a frequency-domain loss that explicitly preserves details.

For a video latent z ∈ RC×T×H×W , where C is the channel dimension, T is the temporal dimension, and H × W are spatial dimensions, we apply a three-dimensional Fast Fourier Transform (FFT), F(z), to transform it into the frequency domain. To extract high-frequency components, we design a 3D high-pass filter:

where σt and σs are cutoff frequencies for temporal and spatial dimensions respectively, and α > 0 controls the steepness of the filter transition.

The high-frequency components are obtained by elementwise multiplication in the frequency domain:

where F -1 denotes the inverse FFT and ⊙ represents element-wise multiplication.

The high-frequency loss measures the difference between student and teacher outputs in the high-frequency domain:

where zθ is the student's output and z∗ is the teacher-refined target.

Algorithm 1 Guided Progressive Distillation (GPD)

Require: Teacher model vϕ, Total steps N, Time steps

{ti}Ni=0 , Student step size K, Text dataset C Ensure: Fast student model vθK 1: Initialize: vθ1 ← vϕ 2: for stage k = 2 to K do 3: Initialize student vθk from vθk-1 4: while not converged do 5: Sample noise ztN ∼ N(0, I) and prompt c from C 6: for i = N down to k step -k do 7: {1. Online Refinement} 8: zti-k+1 ← Sθk-1 (zti , ti, ti-k+1, c) 9: z∗ ti-k ← Sϕ(zti-k+1 , ti-k+1, ti-k, c) 10: {2. Compute Velocity Loss} 11: vtarget ← (z∗ ti-k - zti )/(ti-k - ti) 12: Lv ← ∥vθk (zti , ti, c) - vtarget∥2 13: {3. Conditional High-Frequency Loss} 14: Ltotal ← Lv 15: if k = K and i ≤ 0.5N then 16: zpred ← Sθk (zti , ti, ti-k, c) 17: Lhf ← ∥Fhigh(zpred) - Fhigh(z∗ ti-k )∥2 18: Ltotal ← Ltotal + λ · Lhf 19: end if 20: Update θk ← θk - η∇θk Ltotal 21: zti-k ← z∗ ti-k 22: end for 23: end while 24: end for 25: return vθK

We observe that the generation process exhibits a timedependent behavior in frequency information reconstruction. As shown in Figure 3, the high-frequency components initially resemble unstructured noise during the early stages of the diffusion process. As the diffusion timestep t decreases, the video gradually becomes clearer, and the high-frequency components begin to reveal meaningful structures such as edges and motion details. Based on this observation, we introduce a time-aware high-frequency loss that varies with the diffusion timestep t to better align with this hierarchy.

where K is the final training stage, T is the total diffusion time, and λ0 is the base weight.

The final objective combines trajectory learning with detail preservation:

This joint optimization ensures robust trajectories through online refinement while maintaining high-frequency details. Algorithm 1 shows the complete Guided Progressive Distillation procedure.

on textual prompts, thereby eliminating the overhead associated with pre-synthesizing video data (as required by PeRFlow) or preprocessing large-scale web videos (as required by CausVid). Second, the proposed progressive distillation strategy ensures rapid convergence. As mentioned in the setup, each stage requires only 150 iterations, which substantially reduces the overall training duration.

5.3 Evaluation on Text-to-Video Generation

We evaluate our method on VBench [Huang et al., 2023], a comprehensive video generation benchmark comprising 16 metrics designed to systematically assess both video quality and semantic alignment.

We compare our GPD model with both standard text-tovideo generation baselines and fast-generation approaches. Among traditional full-step generative models, we evaluate against CogVideoX [Hong et al., 2022; Yang et al., 2025], OpenSora [Zheng et al., 2024; Peng et al., 2025], and HunyuanVideo [Kong et al., 2025], all of which represent state-of-the-art transformer-based methods. For acceleration baselines, we include T2V-Turbo-V2 [Li et al., 2024], AccVideo [Zhang et al., 2025], PeRFlow [Yan et al., 2024] and CausVid [Yin et al., 2025]. In particular, PeRFlow and CausVid employ the same base model, training video resolution, and frame count as our method. For our GPD model, we generate videos at two resolutions, 480p and 720p, with a fixed frame length of 81 in both cases.

5 Experiment

5.1 Experiment Setup

We build GPD acceleration pipeline upon Wan2.1-1.3B diffusion architecture [Wan et al., 2025]. Both teacher and student models adopt Wan2.1-1.3B architectures and are initialized with the officially released checkpoint.

For the training configuration, we utilize the textual descriptions from the OpenSora dataset [Zheng et al., 2024; Peng et al., 2025]. The generated video resolution is set to 480 × 832 with 81 frames. Training is conducted on 4 NVIDIA A100 (80GB) GPUs with a batch size of 4. We employ the AdamW optimizer with a learning rate of 1e-6 and a weight decay of 1e-2 . During the progressive distillation process, each stage is trained for 150 iterations. For velocity alignment training, following previous works, we apply classifier-free guidance (CFG) [Ho and Salimans, 2022; Luo et al., 2023a] during the prediction of both the teacher and the frozen student models. The CFG scale is linearly decayed from 6.0 to 1.5 as the training stages progress. Regarding the high-frequency loss, the cutoff frequencies are set to σt = σs = 0.25, and λ0 = 0.5.

Besides, we trained the Wan2.1-1.3B model using the PeRFlow and CausVid methods for comparison. These experiments were conducted on the same 4 NVIDIA A100 (80GB) GPUs, adhering to the hyperparameter settings declared in their original papers. Regarding the training data, CausVid utilizes real-world web videos. In contrast, PeRFlow employs synthetic data generated by the teacher model, for which we utilized the same prompt dataset as our method.

5.2 Comparison on training efficiency

Table 2 presents a quantitative comparison of the training costs across different methods. As shown, our GPD method achieves the lowest computational cost, requiring significantly fewer GPU-days than both PeRFlow and CausVid.

This superior efficiency is attributed to two key factors. First, our training approach is data-efficient; it relies solely Compared to the base Wan2.1-1.3B model, we achieve competitive performance while improving inference efficiency by 8×, significantly enhancing practical usability. Our method completely outperforms the PeRFlow model, as is clearly shown in Table 1. While CausVid operates with fewer inference steps, it suffers from degraded text-alignment and loses fine-grained elements in the generated videos. In contrast, our method maintains strong semantic coherence, as shown in Figure 4.

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.

Our method demonstrates strong generalization across different resolutions; models distilled at lower resolutions still maintain high quality when generating high-resolution videos. The slight decline in the total score for 720p video generation compared to 480p is attributed to the base model, Wan2.1-T2V-1.3B, lacking strong native support for 720p videos.

6 Ablation Studies

6.1 Effect of Online Teacher Guidance

We conduct an ablation study to evaluate the effectiveness of teacher guidance. We first train the student model with constant targets derived from the teacher trajectory; after 3,000 training steps, the model converges and reaches its highest total score. Starting from this checkpoint, we branch into two runs: (i) continuing with constant-target training, and (ii) switching to teacher-refined (GPD) targets. All other training settings are kept identical, and only the training targets differ. Figure 5 reports the evolution of VBench total score together Quality Score

84.4

84.3

84.2

84.1

84.0

83.9

83.8

83.7

3000 3100 3200 3300 3400 3500

Training Iterations

Constant target training

3000 3100 3200 3300 3400 3500

Training Iterations

GPD target training

Table 3: Performance across different inference steps.

with several representative sub-metrics from 3,000 to 3,500 iterations.

Overall, the GPD-target branch consistently outperforms the constant-target branch, reaching its highest peak total score around 3,300 iterations. It also improves quality and aesthetics and, most notably, better preserves motion dynamics than constant targets. These results suggest that teacherrefined targets provide more informative and corrective supervision than static constant targets.

6.2 Progressive Training Analysis

Progressive training is essential for managing the complexity of trajectory learning as step size increases. By gradually increasing the step size across multiple stages, our progressive approach allows the model to incrementally learn longer predictions while maintaining stable performance at each stage.

Table 3 demonstrates the effectiveness of our progressive training strategy across different inference step counts. The results show that our method maintains or even improves performance as the number of inference steps decreases from 24 to 6. Notably, the semantic score consistently improves from 75.98 to 77.99, while the quality score remains stable around 85.5-85.9. The optimal performance is achieved at 6-8 steps, where the model strikes the best balance between generation speed and output quality.

6.3 High-Frequency Loss Analysis

High-frequency components are critical for capturing motion dynamics and preserving visual quality in videos. We compared three high-frequency loss configurations for training GPD: no high-frequency loss, constant high-frequency loss, and weighted high-frequency loss. The model trained without high-frequency loss yielded the lowest dynamic degree and image quality, primarily due to a lack of alignment with highfrequency content. The constant high-frequency loss, which employs a fixed weight across diffusion timesteps, achieved improved scores.

Our weighted high-frequency loss applies a timestepadapted high-frequency loss, dynamically adjusting the loss weight according to the diffusion timestep, as described in Section 4.2. To avoid penalizing unstructured noise in the early denoising stage and focus on the phase where highfrequency components carry meaningful structural information, we restrict the loss to the final distillation stage and lower timesteps (t ≤ 0.5T). As shown in Table 4, this strategy achieves the best performance on both dynamic degree and image quality. We attribute the improvement to the fact that high-frequency reconstruction is highly sensitive to diffusion timesteps; the proposed weighting scheme better matches the stage-dependent nature of the high-frequency formation process.

7 Conclusion

We introduced Guided Progressive Distillation (GPD), a novel framework for accelerating video diffusion models while maintaining high-quality outputs. By combining online refinement, progressive distillation, and high-frequency preservation, GPD achieves an 8× speedup, reducing video generation from 48 steps to just 6 steps, while outperforming existing methods on the VBench benchmark with a total score of 84.04%. Our approach eliminates the need for pre-collected datasets, ensures robust trajectory learning, and preserves fine-grained details, offering a practical solution for efficient, high-quality video synthesis.

Aesthetic Quality

69.8

69.7

69.6

69.5

69.4

69.3

69.2

Dynamic Degree

58

56

54

52

50

3000 3100 3200 3300 3400 3500

Training Iterations

Acknowledgments

This work was supported in part by "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No.2024C01142) and the General Program of National Natural Science Foundation of China (62372403).