The recent advent of powerful video generation models, such as Hunyuan, WanX, Veo3, and Kling, has inaugurated a new era in the field. However, the practical deployment of these models is severely impeded by their substantial computational overhead, which stems from enormous parameter counts and the iterative, multi-step sampling process required during inference. Prior research on accelerating generative models has predominantly followed two distinct trajectories: reducing the number of sampling steps (e.g., LCM, DMD, and MagicDistillation) or compressing the model size for more efficient inference (e.g., ICMD). The potential of simultaneously compressing both to create a fast and lightweight model remains an unexplored avenue. In this paper, we propose FastLightGen, an algorithm that transforms large, computationally expensive models into fast, lightweight counterparts. The core idea is to construct an optimal teacher model, one engineered to maximize student performance, within a synergistic framework for distilling both model size and inference steps. Our extensive experiments on HunyuanVideo-ATI2V and WanX-TI2V reveal that a generator using 4-step sampling and 30% parameter pruning achieves optimal visual quality under a constrained inference budget. Furthermore, FastLightGen consistently outperforms all competing methods, establishing a new state-of-the-art in efficient video generation.
Figures & tables
Figure 1: The model performance is calculated as the average of Vbench’s dynamic degree, aesthetic, and image quality metrics, plotted against sampling steps and parameter retention rate. The parameter-sampling steps trade-offs in FastLightGen compression reveal that performance is initially more sensitive to a reduction in sampling steps than to parameter pruning. As the overall compression ratio increases, the performance degradation from either method begins to converge. Notably, the model with only 30% of its parameters at 4 steps performs on par with the 100% parameter model at 1.2 steps. We report that a moderate compression strategy, specifically, retaining 70% of parameters at 4 sampling steps—strikes a highly effective trade-off, achieving a theoretical speedup of ≈ 35.71 × over the 50-step unpruned baseline which relies on classifier-free guidance (CFG) [ 12 ] .
Figure 2: Vbench average score vs. time spent. MgD: MagicDistillation. Among the accelerated sampling algorithms evaluated, our proposed FastLightGen achieves the greatest speedup, while its performance also surpasses that of the teacher model (Euler).
Figure 3: Overview of three-stage distillation pipeline FastLightGen . FastLightGen begins by identifying less critical layers within the model. The second stage introduces dynamic probabilistic pruning, where these identified layers are stochastically skipped during training. This process yields a robust, stochastically-pruned student model for the final stage. In this final stage, we perform distribution matching. Our "well-guided teacher guidance", which is constructed from the stochastically-pruned student, ensures that the resulting lightweight, few-step generator maintains high performance.
Figure 4: In the first phase of FastLightGen, we identify non-critical layers by visualizing their importance, where a higher error value corresponds to greater layer criticality. For both HunyuanVideo-TI2V and WanX-TI2V, the results reveal a consistent pattern: the initial and final layers are the most critical, while the importance of the intermediate layers is substantially lower.
Figure 5: Visualization of FastLightGen (i.e., 4-step generator that retains 70% of the parameters) across diverse scenarios, including landscapes, food vlogging, dance, and daily activities. Our model generates high-fidelity videos characterized by realistic character motion, detailed expressions, and strong temporal dynamics.
Method
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
average
CogVideoX-I2V-SAT [ 48 ]
0.977
0.955
0.984
0.365
0.598
0.676
0.759
I2Vgen-XL [ 54 ]
0.975
0.964
0.983
0.250
0.653
0.699
0.753
SEINE-512x320 [ 5 ]
0.966
0.942
0.967
0.343
0.584
0.710
0.751
VideoCrafter-I2V [ 4 ]
0.912
0.979
0.980
0.226
0.608
0.717
0.736
SVD-XT-1.0 [ 2 ]
0.975
0.955
0.981
0.524
0.602
0.698
0.789
Step-Video-TI2V [ 27 ]
0.955
0.960
0.992
0.488
0.630
0.704
0.788
Table 1: TI2V quantitative results comparison between FastLightGen and other popular large-scale VDMs on general I2V-VBench. Experimental results for the other comparison methods were directly extracted from the official VBench website. FastLightGen outperforms all prior open-source state-of-the-art image-to-video models, including the distilled teacher, Wan-TI2V.
Method
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
average
time spent
Euler
0.961
0.970
0.982
0.461
0.653
0.711
0.790
885.3s
DMD2 [ 49 ]
0.948
0.946
0.977
0.160
0.583
0.683
0.716
35.4s
LCM [ 36 ]
0.945
0.942
0.979
0.003
0.570
0.665
0.684
35.4s
MagicDistillation [ 33 ]
0.956
0.958
0.980
0.561
0.634
0.701
0.798
35.4s
ICMD [ 44 ]
0.951
0.949
0.982
0.363
0.653
0.715
0.769
396.2s
F3-Pruning [ 39 ]
0.938
0.935
0.977
0.300
0.651
0.706
0.751
396.2s
Table 2: Quantitative comparison of FastLightGen against competing acceleration algorithms on WanX-TI2V. For all metrics, higher is better (↑), except for time spent, where lower is better (↓). FastLightGen outperforms DMD2, LCM, F3-Pruning and ICMD in both speed and performance. FastLightGen achieves an empirical speedup of approximately 35 × over the Euler baseline without performance degradation.
distill weight ( α )
training weight ( 1−α )
action score
vision score
0.0
1.0
2.86
3.41
0.3
0.7
2.88
3.52
0.5
0.5
2.85
3.55
0.7
0.3
2.89
3.58
1.0
0.0
2.92
3.61
Table 3: User study on models trained with varying distill and training weights. Users rated models on a 5-point scale (higher is better) for action quality and visual quality. The results indicate that the best overall performance is achieved when the distill weight α is set to 1 and the training weight 1−α is 0.
Figure 6: Well-guided teacher guidance outperforming approaches that rely solely on weak and strong teachers. Specifically, weak teacher supervision may enable student models to mimic the teacher’s outputs, yet this often leads to suboptimal performance. Conversely, overly strong teachers might provide excessively complex supervision signals, hindering effective knowledge transfer and efficient student learning.
distill weight ( α )
training weight ( 1−α )
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
Teacher
N/A
0.959
0.965
0.991
0.309
0.663
0.709
0.983
0.797
0.0
1.0
0.957
0.964
0.990
0.205
0.648
0.711
0.984
0.780
0.3
0.7
0.959
0.969
0.991
0.224
0.653
0.712
0.985
0.785
0.5
0.5
0.959
0.968
0.991
0.244
0.652
0.712
0.983
0.780
0.7
0.3
0.959
0.965
0.990
0.256
0.654
0.712
0.983
0.788
1.0
0.0
0.961
0.970
0.991
0.268
0.655
0.711
0.984
0.791
Table 4: Quantitative Comparison of FastLightGen under Varying Distill and Training Weights (Teacher: WanX-TI2V). The optimal configuration, with the distill weight set to 1 and the training weight set to 0, enables this setting to outperform all other hyperparameter settings across key metrics, including i2v subject, subject consistency, motion smoothness, dynamic degree, and aesthetic quality.
Inter CFG ( β1 )
Intra CFG ( β2 )
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
Fixed β1=3.5 :
3.5
0.00
0.952
0.956
0.983
0.459
0.646
0.710
0.978
0.812
3.5
0.25
0.952
0.956
0.983
0.500
0.656
0.717
0.975
0.820
3.5
0.50
0.957
0.959
0.986
0.518
0.638
0.717
0.976
0.822
3.5
0.75
0.927
0.886
0.961
1.000
0.584
0.666
0.918
0.848
3.5
1.00
0.909
0.876
0.945
1.000
0.572
0.652
0.909
0.838
Table 5: Quantitative results comparison of β1 and β2 on WanX-TI2V. Fixing β2 at 0.25 yields the best overall performance when β1 is set β1 to 3.5. While β2 values of 0.5, 0.75, and 1.0 achieve higher dynamic degree metrics when β1 is fixed at 3.5, the synthesized videos suffer from noticeable jitter and motions that violate physical laws upon visual inspection. Thus, we select β2=0.25 as the optimal choice.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Stage III
sampling steps
pruning ratio
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
✗
28
30%
0.961
0.970
0.991
0.268
0.655
0.711
0.984
0.791
✓
4
30%
0.960
0.966
0.987
0.378
0.658
0.712
0.977
0.805
✗
28
50%
0.958
0.964
0.988
0.174
0.632
0.692
0.989
0.771
✓
4
50%
0.957
0.958
0.971
0.273
0.649
0.701
0.976
0.784
✗
28
70%
0.959
0.959
0.965
0.064
0.552
0.591
0.983
0.725
✓
4
70%
0.955
0.958
0.953
0.132
0.579
0.619
0.970
0.738
Appendix
Table 6: Quantitative comparison of Stage III application and varying pruning ratios on WanX-TI2V. Applying Stage III, which performs coordinated distillation of sampling step and model size, results in slight declines across metrics like i2v subject, subject consistency, motion smoothness, and temporal flickering. Crucially, however, it achieves significant improvements in dynamic degree, aesthetic quality, and imaging quality. We also observe that larger pruning ratios correlate with poorer model performance overall.
Pruning Ratio
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
0%
0.965
0.984
0.993
0.696
0.603
0.722
0.982
0.849
30%
0.966
0.972
0.996
0.622
0.571
0.693
0.982
0.829
50%
0.948
0.979
0.996
0.501
0.538
0.553
0.977
0.785
70%
0.949
0.967
0.993
0.249
0.465
0.519
0.952
0.728
Appendix
Table 7: Quantitative results for different pruning ratios on HunyuanVideo-ATI2V. As the pruning ratio increases, we observe a significant degradation in model performance, specifically in dynamic degree, aesthetic quality, and imaging quality.
Inter CFG ( β1 )
Intra CFG ( β2 )
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
lipsyncing
Fixed β1=2.0 :
2.0
0.00
0.947
0.941
0.992
0.667
0.585
0.687
0.986
0.829
1.642
2.0
0.25
0.962
0.967
0.994
0.512
0.586
0.702
0.991
0.816
1.967
2.0
0.50
0.962
0.962
0.994
0.217
0.584
0.693
0.990
0.771
1.702
2.0
1.00
0.966
0.970
0.996
0.000
0.587
0.704
0.995
0.745
1.670
Fixed β2=0.25 :
Appendix
Table 8: Quantitative results comparison of β1 and β2 on HunyuanVideo-ATI2V. Fixing β2 at 0.25 yields the best overall performance when β1 is set β1 to 2.0. With β1 fixed at 2.0, we found that higher values for β2 (e.g., 0.5,0.75,1.0) improved aesthetic and image quality but significantly degraded lipsyncing. Therefore, we select β2 =0.25 as it provides the best trade-off between these two objectives.
Teacher Setting
i2v subject
subject consistency
motion smoothness
dynamic degree
aesthetic quality
imaging quality
temporal flickering
average
Pruned
0.947
0.941
0.992
0.667
0.585
0.687
0.986
0.829
Unpruned
0.956
0.932
0.989
0.670
0.525
0.612
0.980
0.809
Appendix
Table 9: Quantitative results on HunyuanVideo-ATI2V. This ablation investigates the choice of teacher model for the real DiT and the fake DiT in Stage III. We find that employing a pruned model as the teacher significantly outperforms using the unpruned version. This result suggests that the optimal teacher is not necessarily the most powerful, but rather one whose capacity is better matched to the student.
Figure 7: In WanX-TI2V, the β2 hyperparameter modulates the dynamic quality of the generated video. A value of β2=0 yields overly static results, whereas β2=1 introduces abrupt temporal discontinuities and excessive visual artifacts.
Figure 8: Additional visualization of FastLightGen (i.e., 4-step generator that retains 70% of the parameters).
Figure 9: Additional visualization of FastLightGen (i.e., 4-step generator that retains 70% of the parameters).
Figure 10: Additional visualization of FastLightGen (i.e., 4-step generator that retains 70% of the parameters).
Diffusion models have achieved impressive performance in video generation, but their iterative denoising process remains computationally expensive due to the large number of tokens processed at each timestep. Recently, progressive resolution sampling has emerged as a promising acceleration approach by reducing latent resolution in early stages. However, scaling this idea to video generation remains challenging, as the additional temporal dimension introduces diverse spatio-temporal demands across different videos, and compressing only a single dimension often leads to limited acceleration or degraded quality. Therefore, we propose DVG, a Dynamic Video Generation framework that jointly allocates computation across time and space, automatically selecting content-aware acceleration strategies without manual tuning or retraining. DVG achieves near-lossless acceleration across models and tasks, reaching up to 7 times speedup on HunyuanVideo and HunyuanVideo-1.5, and 18 times when combined with distillation, demonstrating its potential as a key component in today's large-scale efficient video generation systems. Our code is available at https://github.com/DarrenZheng303/DVG .
Shikang Zheng, Jingkai Huang, Jiacheng Liu +5
Shanghai Jiao Tong University · South China University of Technology · Tsinghua University
Video Diffusion Models (VDMs) have demonstrated superior generation quality but suffer from prohibitive computational costs. While recent few-step distillation techniques significantly accelerate inference, they typically enforce a static model architecture across all denoising stages, ignoring the varying computational demands inherent to different noise levels. In this work, we propose a novel post-training acceleration framework that exploits this redundancy by integrating dynamic structural sparsification directly into the distillation process. Unlike conventional post-hoc compression applied to a fixed diffusion pipeline, our approach jointly optimizes the denoising steps and structured model sparsity, transforming a pre-trained VDM into a compact, step-specific Mixture-of-Models (MoM). To address the training instability arising from this joint optimization, we introduce a Progressive Training Strategy coupled with an Output Rollout Mechanism, which ensures the coherent learning of structural decisions across timesteps. Furthermore, we develop a specialized inference engine to deploy the resulting MoM efficiently. Our method is orthogonal to existing acceleration techniques and highly effective: On Wan-14B, it removes 24% of the per-step FLOPs on top of 4-step distillation, adding a 1.2x wall-clock gain and reaching a 30x speedup over the 50-step teacher while preserving competitive generation quality.
Yu Cheng, Siyue Yao, Zhongang Qi +3
Zhejiang University · BlueImage, vivo · Westlake University
High computational costs and slow inference hinder the practical application of video generation models. While prior works accelerate the generation process through feature caching, they often suffer from notable quality degradation. In this work, we reveal that this issue arises from their inability to distinguish truly redundant features, which leads to the unintended skipping of computations on important features. To address this, we propose PreciseCache, a plug-and-play framework that precisely detects and skips truly redundant computations, thereby accelerating inference without sacrificing quality. Specifically, PreciseCache contains two components: LFCache for step-wise caching and BlockCache for block-wise caching. For LFCache, we compute the Low-Frequency Difference (LFD) between the prediction features of the current step and those from the previous cached step. Empirically, we observe that LFD serves as an effective measure of step-wise redundancy, accurately detecting highly redundant steps whose computation can be skipped through reusing cached features. To further accelerate generation within each non-skipped step, we propose BlockCache, which precisely detects and skips redundant computations at the block level within the network. Extensive experiments on various backbones demonstrate the effectiveness of our PreciseCache, such as achieving an average of 2.6× speedup on Wan2.1-14B without noticeable quality loss.
Jiangshan Wang, Kang Zhao, Jiayi Guo +5
MMLab, CUHK · Tsinghua University · Tongyi Lab, Alibaba