Paper deep dive
Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation
Yuyang You, Yongzhi Li, Jiahui Li, Yadong Mu, Quan Chen, Peng Jiang
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/26/2026, 2:35:01 AM
Summary
The paper introduces a novel distillation framework for video diffusion models to address oversaturation, temporal collapse, and high computational costs. The framework utilizes an adaptive regression loss for spatial supervision, a temporal regularization loss to maintain motion dynamics, and an inference-time frame interpolation strategy to reduce sampling overhead. Experiments on VBench and VBench2 benchmarks demonstrate that the proposed method achieves stable few-step video synthesis with superior perceptual fidelity compared to existing distillation baselines.
Entities (5)
Relation Signals (3)
Adaptive Video Distillation → appliedto → Wan 2.1
confidence 95% · We conduct extensive experiments and ablation studies on the open-source video diffusion model Wan 2.1.
Adaptive Video Distillation → improves → VBench
confidence 95% · Extensive experiments and ablation studies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis
Adaptive Video Distillation → outperforms → Distribution Matching Distillation
confidence 95% · It consistently outperforms existing distillation baselines across multiple metrics.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Video generation has recently emerged as a central task in the field of generative AI. However, the substantial computational cost inherent in video synthesis makes model distillation a critical technique for efficient deployment. Despite its significance, there is a scarcity of methods specifically designed for video diffusion models. Prevailing approaches often directly adapt image distillation techniques, which frequently lead to artifacts such as oversaturation, temporal inconsistency, and mode collapse. To address these challenges, we propose a novel distillation framework tailored specifically for video diffusion models. Its core innovations include: (1) an adaptive regression loss that dynamically adjusts spatial supervision weights to prevent artifacts arising from excessive distribution shifts; (2) a temporal regularization loss to counteract temporal collapse, promoting smooth and physically plausible sampling trajectories; and (3) an inference-time frame interpolation strategy that reduces sampling overhead while preserving perceptual quality. Extensive experiments and ablation studies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion realism. It consistently outperforms existing distillation baselines across multiple metrics.
Tags
Links
- Source: https://arxiv.org/abs/2603.21864v1
- Canonical: https://arxiv.org/abs/2603.21864v1
Trouble viewing inline? Open PDF directly →
Full Text
69,881 characters extracted from source content.
Expand or collapse full text
Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation Yuyang You 1, * Yongzhi Li 2,† Jiahui Li 2 Yadong Mu 1,‡ Quan Chen 2,‡ Peng Jiang 2 1 Peking University 2 Kuaishou Technology yyyou25@stu.pku.edu.cn myd@pku.edu.cn liyongzhi03,lijiahui11,chenquan06,jiangpeng@kuaishou.com Abstract Video generation has recently emerged as a central task in the field of generative AI. However, the substantial compu- tational cost inherent in video synthesis makes model distil- lation a critical technique for efficient deployment. Despite its significance, there is a scarcity of methods specifically designed for video diffusion models. Prevailing approaches often directly adapt image distillation techniques, which fre- quently lead to artifacts such as oversaturation, temporal inconsistency, and mode collapse. To address these chal- lenges, we propose a novel distillation framework tailored specifically for video diffusion models. Its core innova- tions include: (1) an adaptive regression loss that dynami- cally adjusts spatial supervision weights to prevent artifacts arising from excessive distribution shifts; (2) a temporal regularization loss to counteract temporal collapse, pro- moting smooth and physically plausible sampling trajecto- ries; and (3) an inference-time frame interpolation strat- egy that reduces sampling overhead while preserving per- ceptual quality. Extensive experiments and ablation stud- ies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion real- ism. It consistently outperforms existing distillation base- lines across multiple metrics. The source code is publicly available at: https://github.com/yuyangyou/ Adaptive-Video-Distillation 1. Introduction Diffusion models have emerged as the cornerstone of mod- ern generative AI, achieving remarkable progress in both image and video synthesis [1, 2, 8, 14, 16, 20, 41, 43, 50]. Compared with traditional generative paradigms such as GANs [10], diffusion models demonstrate superior per- formance in fidelity, diversity, training stability, and scal- * Work done during an internship at Kuaishou. † Project Leader, ‡ Corresponding authors. Saturation Degree Comparison Dynamic Degree Comparison ModelsModels Saturation Degree Dynamic Degree TeacherDMDrCMOurs TeacherDMDrCMOurs Figure 1. Baselines like DMD and rCM show severe color oversat- uration (left) and reduced motion indicative of temporal collapse (right). Our method achieves appropriate saturation and enhances motion dynamics beyond the teacher model. ability. However, their inference latency remains a ma- jor bottleneck for real-world deployment. This is espe- cially significant in video generation, where existing mod- els often process tens of thousands of tokens simultane- ously in each denoising step and repeat this process for nu- merous iterations, incurring enormous computational over- head. Although training-free acceleration strategies based on specialized samplers [29, 49, 68] can improve efficiency, they still require dozens of sampling steps to produce sat- isfactory results due to the inherent discretization errors of numerical solvers. In contrast, training-based distilla- tion methods enable few-step or even single-step generation while maintaining high-quality outputs. Among existing approaches, Distribution Matching Dis- tillation (DMD) [64, 65] has achieved great success in in- dustrial applications, owing to its strong capability in pre- serving fine-grained image details. Nevertheless, the naive application of DMD usually suffers from oversaturation and mode collapse [28, 64, 65, 70]. These issues become sig- nificantly more severe when extended to video generation (Figure 1). In autoregressive video diffusion models dis- tilled via DMD, temporal error accumulation exacerbates the saturation issue, leading to noticeable quality degrada- tion in long video generation [7, 26, 52]. Meanwhile, mode collapse extends into the temporal dimension, resulting in videos with limited or even static motion. To address these challenges, we introduce a novel dis- arXiv:2603.21864v1 [cs.CV] 23 Mar 2026 tillation framework built upon two key components: an Adaptive Regression Loss and a Temporal Regularization Loss. The Adaptive Regression Loss empowers the student model to more effectively align with the teacher’s distribu- tion, while simultaneously and selectively integrating su- pervision from the real data manifold. This dual-guidance strategy steers the student toward a more optimal target dis- tribution, which effectively resolves the excessive oversatu- ration problem. Concurrently, the Temporal Regularization Loss is specifically designed to counteract the degradation of motion dynamics during distillation. It ensures that tem- poral coherence and motion fidelity are preserved, thereby preventing the temporal mode collapse that leads to static or motion-limited video generation. Crucially, the Adaptive Regression Loss is designed with a dual purpose. Beyond distillation, it intrinsically supports simultaneous supervised fine-tuning, allowing the model to be directly migrated to specialized video distributions, such as anime, advertisements, or visual effects during the distil- lation process. This capability grants our method superior adaptability for practical applications, overcoming the inef- ficiency and performance degradation of traditional meth- ods, which require a sequential process of fine-tuning the teacher model and then performing standard distillation. To further enhance the practicality, we analyze the dif- fusion process and identify that high-noise steps primarily generate high-level semantics, exhibiting minimal tempo- ral feature variance. Based on this insight, we design a novel decoupled temporal interpolation module. This mod- ule strategically performs inference at a lower frame rate during high-noise steps and subsequently interpolates back to the original frame rate in low-noise steps. This optimiza- tion significantly reduces the computational load for infer- ence with negligible impact on generation quality. We conduct extensive experiments and ablation studies on the open-source video diffusion model Wan 2.1. The proposed framework enables the efficient training of few- step student models capable of generating realistic videos with fine spatial details, natural colors, and coherent motion dynamics. On the VBench benchmark, our method con- sistently surpasses state-of-the-art baselines across multiple evaluation dimensions, showcasing both strong quantitative performance and substantial potential for practical deploy- ment. 2. Related Work 2.1. Video Generation and Diffusion Models Early video generation primarily relied on Generative Ad- versarial Networks (GANs) [6, 53, 55], which introduced 3D spatiotemporal convolutions, content-motion disentan- glement, and hierarchical discriminators. More recently, Diffusion Models (DMs) [14, 15, 39, 49] have emerged as the state-of-the-art paradigm, outperforming GANs in training stability, sample diversity, and perceptual fidelity through iterative denoising. A pivotal advancement is La- tent Diffusion Models (LDMs) [43], which perform denois- ing in VAE-projected latent space rather than pixel space, drastically reducing computational overhead while preserv- ing fine-grained spatiotemporal details. Building on this, Flow Matching and Rectified Flow [25, 27] formulate dif- fusion as continuous ODEs with straighter trajectories be- tween noise and data, theoretically enabling fewer sampling steps. State-of-the-art open-source video generation mod- els [1, 2, 4, 5, 48, 54, 56, 63] are predominantly built upon flow or latent-space-based modeling approaches. In paral- lel, Autoregressive (AR) approaches [3, 7, 17, 62, 66, 67] leverage Transformers with causal attention to sequentially generate frames or tokens, supporting variable-length syn- thesis. However, they suffer from error accumulation, lead- ing to temporal inconsistency and motion decay in longer sequences. While significant strides have been made, a crit- ical bottleneck persists: the substantial computational de- mand of leading video generation models. Specifically, dif- fusion models built upon Transformers [40] and classifier- free guidance [13] necessitate tens to hundreds of denois- ing iterations to synthesize even short video segments (5-10 seconds). This computational burden renders them imprac- tical for real-time interactive applications. To address this limitation, our research pivots towards distillation methods, seeking to drastically reduce inference time while maintain- ing the fidelity of the generated content. 2.2. Diffusion Distillation Diffusion model distillation has evolved rapidly, giving rise to multiple complementary paradigms that progressively improve generation efficiency and fidelity. Early studies adopt knowledge distillation [12], which directly super- vises the prediction of noise or score functions to trans- fer the denoising behavior of a pretrained teacher model to a lightweight student. Building upon this idea, progres- sive distillation [23, 38, 44] introduces a recursive frame- work that halves the sampling steps at each stage, allow- ing the student to approximate the teacher’s denoising tra- jectory from coarse to fine. In parallel, adversarial distil- lation [19, 23, 24, 34, 35, 46, 47, 60] explicitly incorpo- rates adversarial objectives to enhance perceptual fidelity and texture sharpness.More recently, consistency distilla- tion [11, 21, 22, 28, 31, 37, 42, 51, 57, 70] reformu- lates diffusion training as a path-matching problem between noisy and clean samples, enabling continuous-time model- ing without relying on discrete solvers. This paradigm pro- motes better mode coverage and can generate videos with high diversity, though often at the expense of fine-detail fi- delity. Meanwhile, score distillation [9, 32, 33, 36, 41, 45, 58, 59, 61, 64, 65, 71] has emerged as a unifying framework for distribution alignment, optimizing the student model Few Step Denoising Generator 퐺 ! Temporal Regularization Loss Adaptive Regression Loss Distribution Matching Loss Real Video Noisy Input Sample Video-Text Pair Sample Text Data Pure Noise Input Text Condition Denoise Compute Regression Loss Loss Mean Cache EMA Update Compute Distribution Matching Loss Gradient Update DMD 푆 !"#$ 푆 %#&" Denoising Output Denoising Output Figure 2. Our method distills a pre-trained teacher model, denoted as s data , into a few-step video generator G φ . The training procedure consists of the following steps:(1) A batch of real video-text pairs is sampled from the dataset. After applying noise perturbations to the videos, the student model performs denoising reconstruction. A regression loss is computed between the reconstructed video and the ground-truth video. Subsequently, this loss is adaptively weighted using our Loss Mean Cache to produce the final adaptive regression loss (see Sec. 3.3 for details). (2) Text conditions are sampled from the dataset to guide the student model in generating a video from pure noise. The denoised output from this process is used to compute a temporal regularization loss (Eq. 8) and a distribution matching loss (Eq. 4).(3) Finally, the generator G φ is updated via gradient descent using the combined losses. The s gen,ξ in DMD are updated separately, following the methodology of DMD2 (this particular update step is not depicted in the figure for clarity). to match the teacher’s score field using statistical distance measures such as reverse KL divergence. A representa- tive example is Distribution Matching Distillation (DMD), which achieves strong fine-grained detail synthesis but suf- fers from mode collapse and over-saturation issues. Sev- eral works attempt to mitigate these problems by combining DMD with adversarial objectives [30, 64] or jointly training it with consistency-based losses [70] in image generation. In contrast, our method builds upon distribution matching distillation, introducing an efficient and stable supervised regularization strategy that effectively addresses these is- sues for video generation. 3. Method In this section, we first analyze the limitations of distribution-matching distillation in video generation (Sec. 3.1). We present a novel training procedure to tackle these challenges, as shown in Figure 2. Its novelty lies in two fundamental components: an adaptive regression loss (Sec. 3.2) that enhances the learning stability and accuracy, and a temporal regularization mechanism (Sec. 3.3) that en- forces temporal consistency across frames. In addition, we present an inference-time frame interpolation acceleration strategy (Sec. 3.4) to enable efficient generation of high- quality videos. 3.1. Analysis of DMD Methods In a standard diffusion process, for samples drawn from the data distribution p(x 0 ), the forward diffusion process is de- fined as x t = α t x 0 + σ t ε, ε∼N (0,I),(1) where x t denotes the noisy sample, α t and σ t are scalar co- efficients that determine the noise schedule and the signal- to-noise ratio at timestep t. The diffusion model learns a reverse denoising process by minimizing the objective L(θ) =E t,x 0 ,ε ∥ε θ (x t ,t)− ε∥ 2 2 ,(2) whereε θ (x t ,t)is a neural network parameterized by θ that predicts the noise at timestep t. Flow-matching mod- els instead predict the instantaneous velocity field, but un- der certain conditions, the two formulations are equivalent. To recover a clean sample x 0 from a noisy initialization x T drawn from a Gaussian prior, the reverse process re- quires not only accurate noise prediction but also numeri- cal solvers to simulate the denoising trajectory. Score dis- tillation leverages pretrained diffusion models as differen- tiable priors to transfer generative capabilities into compact few-step generators. The diffusion model parameterized by θimplicitly defines the data distribution through its score function s θ (x t ,t) =∇ x t logp(x t ) =− 1 σ t ε θ (x t ,t),(3) whereε θ (x t ,t)denotes the noise prediction at timestep t. This score function provides gradients pointing toward high-density regions of the data distribution, serving as a foundation for distillation objectives. Formally, score- based distillation minimizes a divergenceD(p gen,t ∥p data,t ) between the student generator’s marginalp gen,t and the 푆 !"#$ 푆 %#&" Update Direction 푆 !"#$ 푆 %#&" 푆 !"#$ 푆 %#&" 푆 !"#$ 푆 %#&" 푆 !"#$ 푆 %#&" −(푆 !"#$ −푆 %#&" ) '( ! ') Figure 3. This figure explains the origin of oversaturation in distribution-matching distillation. At a given denoising timestep, s real denotes the teacher’s ground-truth clean-sample distribution and s fake the student’s estimate obtained via an online multi-step model; the teacher’s overemphasis on fine-grained detail biases the student toward an oversaturated, suboptimal distribution, de- grading perceptual video quality. teacher diffusion model’s distribution p data,t . Taking DMD as an example, it’s gradient objective can be expressed as ∇ φ L DMD ≜E t ∇ φ KL p gen,t ∥p data,t ≈−E t Z s real Ψ(G φ (ε),t),t − s fake Ψ(G φ (ε),t),t ∂G φ (ε) ∂φ dε , (4) where s real ands fake denote the score functions of the teacher model and an online model trained on the genera- tor’s output distribution, respectively. AndΨ(·)represents the forward diffusion process. Oversaturation Problem. The teacher score provides stable gradients that guide the student generator G φ to- ward matching the target data distribution. However, in practice, training the student model solely with the dis- tribution matching loss often leads to oversaturated visual outputs [64]. As illustrated in Figure 3, this occurs be- cause the teacher model, while encouraging the student to capture finer details, tends to overemphasize local infor- mation, steering the student toward a suboptimal distribu- tion characterized by excessive color saturation. This issue becomes particularly severe in autoregressive generation, where oversaturation can cause error accumulation across frames [7, 17]. To mitigate this, we introduce a direct super- vision mechanism for the student model in Sec. 3.2, effec- t = 0st = 2.5s t = 5s w/o Adaptive Weights w/ Adaptive Weights DMD Teacher Two orange dogs are running. Figure 4. A naive regression loss (Row 3) causes object fusion artifacts (t=2.5s) absent in the teacher (Row 1) and baseline (Row 2). Our adaptive loss (Row 4) resolves this artifact, improving generation quality. tively correcting this bias and alleviating the oversaturation problem in generated videos. TemporalModeCollapse. Moreover,distribution matching losses are known to induce mode collapse in image generation [30, 65, 70], reducing diversity across samples.In video generation, this problem is further amplified by the temporal dimension, manifesting as diminished motion amplitude or even static outputs. Since insufficient motion often degrades perceived video quality more severely than reduced spatial diversity, we introduce a temporal regularization constraint in Sec. 3.3 to explicitly address temporal collapse and enhance motion dynamics in the generated sequences. 3.2. Adaptive Regression Loss Although the teacher model provides strong gradient guid- ance in high-quality regions, the distribution matching dis- tillation effectively transfers the teacher’s ability to synthe- size fine-grained details. However, this detail-oriented opti- mization often leads to oversaturation, especially in autore- gressive generation, where such artifacts accumulate over time. To alleviate this issue, we introduce a regression loss that directly supervises the student model, injecting real data in- formation to jointly guide the gradient updates alongside the distillation loss. Nevertheless, when the student model si- multaneously fits the teacher distribution and is influenced by real samples that deviate significantly from it, the op- timization may converge to a suboptimal intermediate dis- tribution. As illustrated in Figure 4, this can cause visual artifacts such as tearing, unnatural object blending, duplica- tion, or disappearance. To address this problem, we propose an adaptive regres- sion loss, which flexibly regularizes the student output to- ward real data samples while suppressing the influence of inconsistent data points by assigning them lower weights. The loss is defined as L = w t,s ∥ε θ (x t ,t)− ε∥ 2 2 ,(5) wherew t,s is a learnable, timestep-dependent adaptive weight. For real video data sampled at different locations, the deviation between the generated and true distributions varies significantly. Points with larger deviations tend to cause overly aggressive gradient updates, resulting in tear- ing artifacts or suboptimal frames. To alleviate this, we assign lower weights to such points. We maintain a dy- namic, learnable cache to estimate the expected loss value at each timestep tof the student’s few-step inference sched- ule. This cache is updated using an exponential moving average (EMA): ̄ L t,s = αL t,s−1 + (1− α)L s ,(6) whereαdenotes the EMA coefficient,tindexes the diffu- sion timestep, and srepresents the current training iteration. Based on this cache, the adaptive gradient weight for the current sample is computed as ω t,s = 1−σ k· (L s − ̄ L t,s−1 ) , σ(x) = 1 1 + e −x ,(7) wherekis a scaling factor andσ(·)is the Sigmoid func- tion. This adaptive weighting mechanism enables the stu- dent model to smoothly align with the real data distribution while suppressing overexposure. Moreover, it improves both temporal and spatial diversity, effectively alleviating mode collapse that often arises during few-step distillation 3.3. Temporal Regularization As discussed in Sec. 3.1, compared to image generation, the impact of mode collapse on the additional temporal dimension in video generation is non-negligible.This phenomenon often manifests as limited motion dynamics, sometimes even near-static sequences or reduced variations in object appearances (See appendix Sec. 7 for visual ex- amples). In certain application scenarios, such degradation in temporal coherence can be far more detrimental to video quality than the loss of spatial diversity typically caused by mode collapse. The adaptive regression loss introduced in the previous section partially alleviates spatial mode col- lapse by incorporating real data supervision, enabling the student to explore unseen modes. However, its supervision over the temporal dimension remains weak. Compared with reduced object diversity, the lack of motion typically has a far more detrimental effect on perceived video quality. To address this, we introduce a motion regularization loss that explicitly encourages temporal variation in the generated sequences: L temp =− log E x∼p θ [Var(x)] + ε ,(8) wherexdenotes samples generated by the student model, and the varianceVar(x)is computed along the temporal di- mension. The constant εensures numerical stability. This regularizer penalizes degenerate solutions with low tempo- ral variance, effectively promoting motion diversity. To pre- vent it from dominating the optimization once the model es- capes the collapsed regime, the loss is truncated after con- vergence to a sufficiently diverse temporal distribution. 3.4. Frame-interpolation Inference We propose a frame interpolation strategy to accelerate video diffusion model inference, motivated by the oper- ational dichotomy in denoising: high-noise steps capture coarse semantics, while low-noise steps refine details [37, 52]. Considering the correlation between the high-noise de- noising stage and the inter-frame similarity of clean samples (visualized in Figure 15), our approach reduces the frame rate by half during the high-noise stage (e.g., first 2 of 4 steps). A lightweight, pre-trained UNet module then in- terpolates the sequence back to the full frame rate within the VAE’s latent space. This allows the subsequent low- noise denoising to erase interpolation-induced artifacts. The UNet is pre-trained on real data via a regression loss to pre- dict the features of an intermediate frame from its neighbors (appendix Sec. 8). This method substantially cuts computa- tional costs by shortening the effective sequence length for inference, with a negligible impact on perceptual quality. Algorithm 1: Training Procedure Input : Pretrained teacher model s real ; Video dataD =c, y; Few-step denoising timestepsT =0,t 1 ,t 2 ,...,t Q Output: Few-step generator G G,s fake ← init(s real ); ̄ L reg ← 0 while not converged do if is generator update step then / * Update Generator G * / 1x← G(z∼N (0,I), c∼D) 2 L KL ← DMDLoss(s real ,s fake , x) // Eq.(4) 3 L temp ← TemporalLoss(x) // Eq.(8) 4y t ← ForwardDiffusion((c y , y)∼D,t∼T ) 5ˆy ← G(y t , c y ) 6 L reg ← RegressionLoss(ˆy, y) // Eq.(5) 7 ω t,s ← f ( ̄ L reg ,L reg ,t) // Eq.(7) 8 L G ←L KL + ω temp L temp + ω reg ω t,s L reg 9 G← G− η G ∇ G L G 10 ̄ L reg ← (1− α) ̄ L reg + αL reg // EMA update end / * Update online model s fake * / 11x t ′ ← ForwardDiffusion(stop grad(x),t ′ ∼U (0, 1)) 12 L denoise ← DenoisingLoss(s fake (x t ′ ,t ′ ), stopgrad(x)) // Eq.(2) 13 s fake ← s fake − η s ∇ s fake L denoise end Table 1. Quantitative comparison on two benchmarks (VBench2 and VBench1). Our method achieves the best overall score across all metrics and datasets. DMD ∗ is denoted as our baseline method. In all result tables, bold text indicates the optimal result, while underlined text represents the suboptimal result. MethodStepsResolutionParamsTime VBench2 Evaluation DimensionVBench1 Evaluation Dimension Creati- vity Common- sense Controll- ability Human Fidelity PhysicsTotal Quality Score Semantic Score Total Teacher50×2 832×480×81 1.3B270s48.0158.5121.8280.6545.9750.9984.1164.2080.13 DMD ∗ [65]4832×480×81 1.3B10.8s55.9056.5125.2986.7543.7253.6385.5561.1080.66 LCM[31]6832×480×81 1.3B16.2s27.6645.5310.9576.0740.4040.1277.6250.1572.12 PCM[57]6832×480×81 1.3B16.2s 44.5450.4217.0281.4843.5347.4079.3651.4073.70 DCM[37]6832×480×81 1.3B16.2s61.3152.7116.9383.9944.0751.8079.2752.1673.92 rCM[70]4832×480×81 1.3B10.8s 46.4060.7929.1982.9850.8154.0384.8061.5580.15 Ours4832×480×81 1.3B7.8s48.1959.0729.3988.2650.4855.0885.2765.6581.35 Teacher50×2 832×480×8114B730s50.6357.3925.6287.3339.7252.1484.4470.4781.64 DMD ∗ 4832×480×8114B28s54.9259.3133.8683.7852.4856.8782.2269.2779.63 LCM6832×480×8114B42s36.7151.2819.8771.5543.0444.4980.3353.8475.03 DCM6832×480×8114B42s57.2056.8218.9184.4844.8252.4480.8055.5175.73 Ours4832×480×8114B22.2s52.8965.0930.1789.0058.1359.0684.7074.0682.57 3.5. Overall Training Process The overall training pipeline is summarized in Algorithm 1. Both the student generator G and the online model s fake used for computing the distribution-matching loss are ini- tialized with the teacher model’s weights at the beginning of training. We follow the framework of DMD2 [64]. For the distribution matching loss, we only use fake videos gen- erated by G with text conditions. The online model s fake is updated using the two timescale update rule from DMD2. This means that before each update to the student model, we first optimize s fake for multiple steps using videos sam- pled from the student. This helps s fake adapt to the student’s current output distribution. Our temporal regularization loss is calculated on the same videos generated for distribution matching. Therefore, this step adds no extra forward pass cost. The regression loss, however, is calculated using the real video dataset. As shown in Eq. 6, we maintain a sepa- rate exponential moving average (EMA) mean for each de- noising step of the student model to compute the adaptive weights. Overall, our training process adds only one addi- tional forward pass per update to the student model. The final loss function is formulated as: L G =L KL + ω reg ω t,s L reg + ω temp L temp .(9) 4. Experiments 4.1. Experimental Setups We employ Wan2.1-T2V-1.3B and Wan2.1-T2V-14B as the teacher models for our experiments. Starting from their of- ficially released pretrained weights, we distill them into a few-step denoising student model capable of generating 5- second, 16 fps, 832 × 480 resolution high-quality videos. For training data, the distribution matching loss is condi- tioned on text annotations from a mixed dataset of open- source and proprietary videos, without direct use of the video data itself, following the methodology of DMD2 [64]. The regression loss, in contrast, is computed on a high- quality subset of 150,000 video samples, which we curated and cleaned from online sources. Further details on our data pre-processing are available in the appendix Sec. 6. We use the AdamW optimizer with lr = 2 × 10 −6 , an EMA decay factor α = 0.95, and k = 3.0 for adaptive weight computation in Eq. 8. The weights ω reg and ω temp in Eq. 9 are set to 2.0 and 0.05, respectively. The temporal regular- ization loss is truncated once it converges to around 0.6, and the teacher model employs a classifier-free guidance (CFG) scale of 5.0. The hyperparameter settings are discussed in detail in appendix Sec. 9. All experiments are conducted on 24 GPUs. 4.2. Evaluation Results Benchmark Evaluation. We conduct comprehensive quantitative evaluations on both VBench2 [69] and VBench1 [18] benchmarks to compare our approach against several state-of-the-art diffusion distillation methods 1 , in- cluding DMD (baseline). The results are summarized in Table 1. All methods are distilled from the same teacher model, using its officially released pretrained weights, and evaluated with the inference steps consistent with their re- spective training configurations. Our method achieves state- of-the-art performance on both benchmarks in terms of 1 For methods with open-source weights, we use their released check- points. For others, we retrain the models on the same datasets following the original hyperparameter settings. All models are evaluated in a unified codebase with BF16 precision to ensure fair and consistent comparisons. DMD DCM rCM Ours t = 0st = 2.5st = 5s t = 0st = 2.5st = 5s Thevideoopenswithasereneforestscene,showcasingacarpetofgreengrasswithbrownleavesandsmalltwigs,bathedinsunlight filteringthroughthetrees,creatingatranquilatmosphere.Atthetwo-secondmark,apersonwearinglightbrownhikingbootsanddarkgreen socksisseencrouchedonthegrass,tyingtheirshoelacesinaforestclearing,surroundedbytalltreesanddryleaves,suggestingapeaceful outdoorsetting.Thesamescenecontinuesatthefour-secondmark,emphasizingtheperson'ssolitaryactivityinthenaturalenvironment. Amanwithcurlyhairandabeard,wearingapatternedshirt,standsinadomestickitchenwithtiledwalls,awirestrainer,shelvesofcanned food,andaredornament.Hegazesintensely,brieflydistractedbutmaintainingafocused,intenseexpression.Thebackgroundremains unchangedthroughout. Figure 5. Qualitative comparison. Baseline methods tend to produce oversaturated colors (left case) and exhibit stiff or static motion (right case), while our method generates videos with more natural color tones, smoother temporal dynamics, and scene transitions that better align with the prompt. The highlighted text in the prompt indicates key visual changes described in the scene. Zoom in for better visualization. the Total Score. Specifically, on VBench2, our approach shows substantial improvements in Human Fidelity com- pared to existing methods, demonstrating more precise mo- tion alignment and perceptual realism. On VBench1, our model attains the highest Semantic Quality score, further confirming its superior generalization and generation fi- delity across diverse prompts. Furthermore, we obtained the highest total scores on both models and benchmarks. Qualitative Comparisons. We further evaluate the gen- eralization capability of our distilled model by assessing its performance on high-quality prompts unseen during train- ing. As illustrated in Figure 5, our method demonstrates a significant reduction in color oversaturation while produc- ing richer fine-grained details and smoother, more extensive motion. These improvements markedly enhance the per- ceptual realism of videos generated from complex, detail- rich prompts, offering a distinct advantage in human visual perception. Furthermore, we investigate the adaptability of our method for downstream tasks through supervised fine- tuning, facilitated by the adaptive regression loss during dis- tillation. As shown in Figure 6, by fine-tuning on a small- scale animation dataset, our method achieves remarkable style transfer. The student model becomes capable of gen- erating videos in a specific animation style that the original teacher model fails to produce. This result strongly vali- dates the practical utility and adaptability of our approach. User Study. We conducted a human preference study to evaluate generation quality. Long-form prompts were cre- ated using a large language model, and from these prompts we produced a total of 180 comparison video samples. AnanimatedclipfromPleasantGoatandBigBigWolf.Twocartoonsheep,onewhite andonebrown,standsidebysideinalushgreenforest.Theyhavethickwool,distinctive horns,andexpressivefaceswithlargeeyesandsmallmouths.Theirexpressionsshift fromsurprisetoconfusiontoshock,indicatingunexpectedeventsinthepeacefulforest. t = 0s t = 2.5st = 5s DMD Teacher Ours GT Figure 6. The figure demonstrates that the adaptive-regression-loss enables effective distribution transfer during distillation. ”GT” de- notes ground-truth frames; both the Teacher and DMD models fail to generate specific anime-style videos consistent with the target domain. We recruited 12 professional, independent annotators; each paired sample was evaluated by at least three distinct an- notators. The annotators were instructed to select the bet- ter video based on two criteria: visual quality and seman- tic alignment with the input prompt, with an option to rate them as comparable. For each comparison, we aggregated judgments by majority vote to produce the final preference. As shown in Figure 7, our model consistently outperforms Figure 7. The figure presents a user study comparing our dis- tilled student model against its teacher (Wan2.1) and other existing video diffusion models. The results demonstrate that our model achieves superior video quality, even surpassing its teacher model in terms of user preference, while operating at a significantly lower inference cost. all baseline methods and is even preferred by users over its teacher model. This strongly demonstrates that our model achieves a more user-favored generation quality while op- erating at a much faster inference speed. 4.3. Efficiency Analysis One single denoising step of the 1.3B model takes 2.7 sec- onds on one GPU, while inference at half frame rate reduces this time to 1.1 seconds. To further enhance efficiency, we employ a lightweight UNet-based interpolation network to upsample VAE-encoded video sequences. Specifically, our 4-step student model performs the first two denoising steps at half frame rate, and before the third denoising step, the interpolation network restores the sequence to its original frame rate. This interpolation process is extremely fast, tak- ing only a few hundred milliseconds.This inference design effectively balances video quality and computational effi- ciency, achieving a 30% acceleration in overall inference speed without compromising perceptual fidelity. 4.4. Ablation Studies To validate the effectiveness of the proposed mechanisms, we conduct a series of evaluations based on the 1.3B DMD baseline. Specifically, we compare models trained with (1) temporal regularization, (2) regression loss, and (3) adap- tive regression loss. Adaptive Regression Loss. The Instance Preservation metric measures the temporal consistency of objects within a video. A lower score indicates instability, such as object merging, splitting, sudden appearance, or disappearance. As shown in Table 2, directly incorporating the regression loss leads to a significant drop in the Instance Preservation score compared to the model without it, indicating that the Table 2. Ablation study on the effectiveness of the proposed Adap- tive Regression Loss and Temporal Regularization. DMD serves as the baseline method, while TR and AdaLoss denote models equipped with the Temporal Regularization and Adaptive Regres- sion Loss, respectively. Full+VIF refers to the model incorporat- ing both components and frame-interpolation. MethodStepsResolutionTime Evaluation Score Instance Preservation Dynamic Degree Teacher 50× 2 832× 480270s92.3985.56 DMD4 832× 48010.8s88.8872.22 +TR4 832× 48010.8s85.38100.00 +TR+RegLoss 4 832× 48010.8s83.0478.61 +TR+AdaLoss 4 832× 48010.8s92.3999.72 Full+VIF4 832× 4807.8s91.8197.77 student model tends to produce more artifacts, such as tear- ing and object distortion. The examples in Figure 4 also illustrate this point. When replaced with the adaptive re- gression loss, the score improves markedly, even surpassing the DMD baseline and approaching the performance of the teacher model. This improvement stems from the model’s ability to focus on learning in reliable data regions while mitigating hallucinations in poorly aligned samples. Temporal Regularization. The Dynamic Degree metric evaluates the overall motion dynamics by estimating the optical flow magnitude across frames, where the score re- flects the proportion of videos exhibiting meaningful mo- tion among all test samples. As shown in Table 2, with- out the temporal regularization loss, DMD-based distilla- tion suffers from a severe degradation in Dynamic Degree (over 10 percentage points). Introducing temporal regular- ization effectively restores motion diversity, ensuring that nearly all generated videos exhibit meaningful dynamics. Furthermore, when combined with the regression loss, the model maintains robustness and preserves a high level of motion fidelity. 5. Conclusion In this work, we addressed the key challenges of oversatura- tion and temporal inconsistency in the distillation of video diffusion models. We propose a novel framework that in- corporates an adaptive regression loss to prevent spatial ar- tifacts and correct color bias, alongside a temporal regu- larization loss to preserve motion dynamics and mitigate mode collapse. Combined with an inference-time accelera- tion strategy, our method enables stable, few-step (4-step), and fast video synthesis. Extensive experiments demon- strate that our approach significantly outperforms existing baselines on two VBench benchmarks, achieving superior perceptual fidelity and motion realism. References [1] Fan Bao, Chendong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu.Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024. 1, 2 [2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2 [3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Ranjan, and Bj ̈ orn Ommer. Stable video diffusion: Scaling autoregressive models for video generation. 2024. 2 [4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1(8):1, 2024. 2 [5] Zixu Chen, Jianfeng Wang, Gengwei Liu, Xiaoxu Liu, Jun Liu, Yuan Liu, Lewei Liu, Hongsheng Li, and Hong Wang. Videocrafter2: Over-1-minute video generation with diffu- sion models. arXiv preprint arXiv:2401.09047, 2024. 2 [6] Aidan Clark, Jeff Donahue, and Karen Simonyan.Effi- cient video generation on complex datasets. arXiv preprint arXiv:1907.06571, 2019. 2 [7] Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, and Cho-Jui Hsieh. Self- forcing++: Towards minute-scale high-quality video genera- tion. arXiv preprint arXiv:2510.02283, 2025. 1, 2, 4 [8] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7346–7356, 2023. 1 [9] Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu, Fanzhou Wang, Zhiqian Lin, Tianxiang Ren, Dahua Lin, Ruihao Gong, and Lei Yang.Phased dmd: Few-step distribution matching distillation via score matching within subintervals. arXiv preprint arXiv:2510.27684, 2025. 2 [10] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 [11] Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Mul- tistep consistency models. arXiv preprint arXiv:2403.06807, 2024. 2 [12] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean.Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. 2 [13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 [14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2 [15] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Rui Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Video diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 [16] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion mod- els. arXiv preprint arXiv:2210.02303, 2022. 1 [17] Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman.Self forcing: Bridging the train- test gap in autoregressive video diffusion. arXiv preprint arXiv:2506.08009, 2025. 2, 4 [18] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Com- prehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2024. 6 [19] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into condi- tional gans. In European Conference on Computer Vision, pages 428–447. Springer, 2024. 2 [20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565–26577, 2022. 1 [21] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Mu- rata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory mod- els: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 2 [22] Sangyun Lee, Yilun Xu, Tomas Geffner, Giulia Fanti, Karsten Kreis, Arash Vahdat, and Weili Nie. Truncated con- sistency models. arXiv preprint arXiv:2410.14895, 2024. 2 [23] Shanchuan Lin, Anran Wang, and Xiao Yang.Sdxl- lightning:Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2 [24] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 2 [25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. arXiv preprint arXiv:2210.02747, 2022. 2 [26] Jinxiu Liu, Xuanming Liu, Kangfu Mei, Yandong Wen, Ming-Hsuan Yang, and Weiyang Liu. Streaming autoregres- sive video generation via diagonal distillation, 2026. 1 [27] Xingchao Liu, Chengyue Gong, and Qiang Liu. Rectified flow: A general ode model for generative modeling. In In- ternational Conference on Learning Representations (ICLR), 2023. 2 [28] Cheng Lu and Yang Song.Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 2 [29] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffu- sion probabilistic model sampling in around 10 steps. Ad- vances in neural information processing systems, 35:5775– 5787, 2022. 1 [30] Yanzuo Lu, Yuxi Ren, Xin Xia, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Andy J Ma, Xiaohua Xie, and Jian-Huang Lai. Adversarial distribution matching for diffusion distil- lation towards efficient image and video synthesis. arXiv preprint arXiv:2507.18569, 2025. 3, 4 [31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao.Latent consistency models: Synthesizing high- resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 6 [32] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffu- sion models. Advances in Neural Information Processing Systems, 36:76525–76546, 2023. 2 [33] Weijian Luo, Zemin Huang, Zhengyang Geng, J Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching. Advances in Neural Information Process- ing Systems, 37:115377–115408, 2024. 2 [34] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to- image synthesis by self-cooperative diffusion gans. arXiv preprint arXiv:2403.12931, 2024. 2 [35] Yihong Luo, Tianyang Hu, Yifan Song, Jiacheng Sun, Zhen- guo Li, and Jing Tang. Adding additional control to one-step diffusion with joint distribution matching. arXiv preprint arXiv:2503.06652, 2025. 2 [36] Yihong Luo, Tianyang Hu, Jiacheng Sun, Yujun Cai, and Jing Tang. Learning few-step diffusion models by trajectory distribution matching.arXiv preprint arXiv:2503.06674, 2025. 2 [37] Zhengyao Lv, Chenyang Si, Tianlin Pan, Zhaoxi Chen, Kwan-Yee K Wong, Yu Qiao, and Ziwei Liu. Dcm: Dual- expert consistency model for efficient and high-quality video generation. arXiv preprint arXiv:2506.03123, 2025. 2, 5, 6 [38] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14297–14306, 2023. 2 [39] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR, 2021. 2 [40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205, 2023. 2 [41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2 [42] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajec- tory segmented consistency model for efficient image syn- thesis. Advances in Neural Information Processing Systems, 37:117340–117362, 2024. 2 [43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ̈ orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2 [44] Tim Salimans and Jonathan Ho.Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 2 [45] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. Advances in Neural Information Process- ing Systems, 37:36046–36070, 2024. 2 [46] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 2 [47] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87–103. Springer, 2024. 2 [48] Team Seawead, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, et al.Seaweed-7b: Cost-effective train- ing of video generation foundation model. arXiv preprint arXiv:2504.08685, 2025. 2 [49] Jiaming Song,Chenlin Meng,and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1, 2 [50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. arXiv preprint arXiv:2011.13456, 2020. 1 [51] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2 [52] Hansi Teng, Hongyu Jia, Lei Sun, Lingzhi Li, Maolin Li, Mingqiu Tang, Shuai Han, Tianning Zhang, WQ Zhang, Weifeng Luo, et al. Magi-1: Autoregressive video genera- tion at scale. arXiv preprint arXiv:2505.13211, 2025. 1, 5 [53] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 [54] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kin- dermans, S. E., D. T., J.A., S.G., D.A., H.A., J.L., D.G., G.P., E.S.H.A., and A.A. Phenaki: Variable length video genera- tion from open domain text. In International Conference on Learning Representations (ICLR), 2023. 2 [55] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. In Advances in Neu- ral Information Processing Systems (NeurIPS), 2016. 2 [56] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 2 [57] Fu-Yun Wang, Zhaoyang Huang, Alexander Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models. Advances in neural information pro- cessing systems, 37:83951–84009, 2024. 2, 6 [58] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion. Advances in neural information processing systems, 36: 8406–8441, 2023. 2 [59] Sirui Xie, Zhisheng Xiao, Diederik Kingma, Tingbo Hou, Ying Nian Wu, Kevin P Murphy, Tim Salimans, Ben Poole, and Ruiqi Gao. Em distillation for one-step diffusion mod- els. Advances in Neural Information Processing Systems, 37: 45073–45104, 2024. 2 [60] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image gener- ation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196–8206, 2024. 2 [61] Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffu- sion models with f -divergence distribution matching. arXiv preprint arXiv:2502.15681, 2025. 2 [62] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Yusuf Aytar. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 [63] Wilson Yan, D. Chen, C.C., M.T., T.M., M.E., Y.C., A.B., P.B., Y.W., X.H., and M.H.C. Moviegen: A high-quality and controllable video generation model with stacked temporal- spatial transformers.arXiv preprint arXiv:2403.02324, 2024. 2 [64] Tianwei Yin, Micha ̈ el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman.Im- proved distribution matching distillation for fast image syn- thesis. Advances in neural information processing systems, 37:47455–47487, 2024. 1, 2, 3, 4, 6 [65] Tianwei Yin, Micha ̈ el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6613–6623, 2024. 1, 2, 4, 6 [66] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Free- man, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion mod- els. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22963–22974, 2025. 2 [67] Lijun Yu, Yong Cheng, Kihyuk Sohn, Jos ́ e Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming- Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: Masked generative video transformer, 2023. 2 [68] Qinsheng Zhang and Yongxin Chen. Fast sampling of dif- fusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 1 [69] Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, Yu Qiao, and Ziwei Liu. Vbench-2.0: Advancing video genera- tion benchmark suite for intrinsic faithfulness. arXiv preprint arXiv:2503.21755, 2025. 6 [70] Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jin- tao Zhang, Yogesh Balaji, Jianfei Chen, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Large scale diffusion distillation via score-regularized continuous-time consistency.arXiv preprint arXiv:2510.08431, 2025. 1, 2, 3, 4, 6 [71] Mingyuan Zhou,Huangjie Zheng,Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In Forty-first International Confer- ence on Machine Learning, 2024. 2 Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation Supplementary Material Contents 6 . Training Data5 6.1 . Data Collection . . . . . . . . . . . . . . .5 6.2 . Data Filtering . . . . . . . . . . . . . . . .5 6.3 . Data Captioning . . . . . . . . . . . . . .5 7 . Quality Visualize5 8 . Frame-interpolation Module6 9 . Hyperparameter Discuss7 9.1 . Temporal Regularization . . . . . . . . . .7 9.2 . Adaptive Regression Loss . . . . . . . . .8 1 DMD rCM Ours An adorable happy otter confidently stands on a surfboard wearing a yellow lifejacket, riding along turquoise tropical waters near lush tropical islands, 3D digital render art style. DMD rCM Ours A close up view of a glass sphere that has a zengarden within it. There is a small dwarf in the sphere who is raking the zengarden and creating patterns in the sand. DMD rCM Ours The camera rotates around a large stack of vintage televisions all showing different programs —1950s sci-fi movies, horror movies, news, static, a 1970s sitcom, etc, set inside a large New York museum gallery. Figure 8. ... DMD rCM Ours A Japanese animated film of a young woman standing on a ship and looking back at camera. DMD rCM Ours A woman singing and standing in a concert stage with a bright light in the background. DMD rCM Ours Panda playing the guitar Figure 9. ... DMD rCM Ours A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about. DMD rCM Ours Animated scene features a close-up of a short fluffy monster kneeling beside a melting red candle. The art style is 3D and realistic, with a focus on lighting and texture. The mood of the painting is one of wonder and curiosity, as the monster gazes at the flame with wide eyes and open mouth. Its pose and expression convey a sense of innocence and playfulness, as if it is exploring the world around it for the first time. The use of warm colors and dramatic lighting further enhances the cozy atmosphere of the image. Figure 10. Comparison with baselines on MovieGen prompts. While DMD and rCM (a strong performer from Sec. 4) produce videos with low motion dynamics and oversaturated colors, our method resolves both issues. It achieves superior motion, well-calibrated colors, and excellent detail and stability.More video examples are included in the supplementary zip file. (Note: We have confirmed that any visual artifacts in some videos are not code-level issues.) 6. Training Data This section details the pipeline for the collection, filtering, and annotation of the video dataset used in our training. 6.1. Data Collection For the construction of a high-fidelity training dataset, we curated video data from two primary channels: (1) publicly available, open-source videos from the internet, and (2) our internal, proprietary archives. To achieve broad diversity and coverage, we deliberately included videos from a mul- titude of categories. These range from everyday scenes like lifestyle, travel, and food, to more specialized domains such as aerial cinematography, medical imaging, and pro- fessional sports, among many others. 6.2. Data Filtering Our data filtering pipeline consists of two main stages de- signed to ensure the quality of the training videos. Preliminary Filtering by Resolution. We first decode each video to determine its spatial resolution. To ensure suf- ficient source quality, we filter out all videos with a width or height below 720p (1280×720). Subsequently, all re- maining videos are uniformly resized to 480p for training, a process that helps preserve essential structural details from the high-resolution source material. Frame-based Quality Assessment. From each remain- ing video, we uniformly sample four frames at fixed tem- poral positions (specifically, the 1st, 21st, 41st, and 61st frames). Each sampled frame must pass two complemen- tary quality checks: • Monochromatic Scene Detection: We convert the frame to the HSV color space and compute the entropy of its hue histogram. If the hue entropy is below a prede- fined threshold of 0.60, the frame is classified as part of a monochromatic scene (e.g., all-black, all-white, or solid- color backgrounds), and the corresponding video is re- jected. • Blur Detection: We assess the sharpness of the frame by calculating the variance of its Laplacian operator. If the Laplacian variance is lower than 20.0, the frame is deemed excessively blurry (potentially due to out-of- focus or motion blur issues), and the video is eliminated. Optical Flow Analysis for Motion Dynamics. To ensure that our training videos contain sufficient dynamic content, we perform an optical flow analysis on all videos that passed the initial screening. We use an optical flow algorithm to extract motion vector fields between consecutive frames and then compute the average motion magnitude. Videos with a magnitude below 0.2 are considered to have insignificant Prompt： “Please describe the details of this video, focusing on concrete aspects such as objects, people, colors, actions, and spatial relationships. Do not include any subjective descriptions related to symbolism, implication, or emotions.” Captionexample： "An elderly man with white hair and a mustache, wearing a red, blue, and white checkered shirt, is seen carefully hanging a large number of vibrant red chili peppers on a wall outdoors, likely in a rural or traditional setting. The scene is bathed in sunlight, highlighting the warmth of the day and the rich colors of the peppers. The man's focused expression and the meticulous way he handles the peppers suggest a deep connection to the cultural practice of preserving food." Figure 11. Caption model prompt and caption style examples. motion (e.g., static scenes or overly stable footage) and are excluded from the training set. Temporal Consistency Filtering. To further enforce temporal coherence within video clips, we introduce an as- sessment based on frame-to-frame consistency. For the re- maining videos, we calculate a temporal consistency score following the methodology of VBench1 [YourVBenchCita- tion]. We retain only the top 50% of videos ranked by this score. Aesthetic Quality Filtering. Finally, we apply an aes- thetic quality filter to ensure visual appeal. We employ a pre-trained visual aesthetic assessment model to score keyframes. Specifically, we uniformly sample 8 frames from each video, calculate their average aesthetic score, and keep only the top 50% of videos based on this metric. 6.3. Data Captioning For the filtered data, we employ a proprietary 7B-parameter captioning model for annotation. The prompt and style ex- amples used for the annotation process are shown in Fig- ure 11. 7. Quality Visualize To further evaluate our method’s performance in practical video generation scenarios, we conducted a qualitative com- parison on the official MovieGen prompt set against the Figure 12. This figure illustrates the impact of the truncation threshold of the temporal regularization loss on model perfor- mance. The horizontal axis represents the truncation threshold of the regularization loss, while the vertical axis shows the motion score and the instance preservation score of the videos generated by the student model, respectively. Figure 13. Visualization of the adaptive weight function in Eq. (7) for different values of k. The x-axis represents the deviationL s − ̄ L s , while the y-axis represents the resulting adaptive weight. DMD baseline and rCM, the latter of which performed fa- vorably in our main experiments (Sec. 4). All methods gen- erated videos under identical prompts and consistent sam- pling configurations to ensure a fair comparison. The re- sults, as illustrated in Figures 8, 9, and 10, demonstrate that our method consistently achieves superior and more sta- ble performance across multiple key dimensions. Specifi- cally, in terms of motion dynamics, our approach gener- ates videos with more extensive action, cinematic camera movements, and greater fluidity compared to the baselines. Regarding color saturation and overall style, our method produces visuals with well-calibrated saturation, avoiding the oversaturation artifacts common in other approaches and resulting in a style that aligns more closely with natural video distributions. Furthermore, concerning detail qual- ity, our model excels at preserving fine-grained textures and sharp structural contours, effectively mitigating issues such as blurring and local structural degradation. 8. Frame-interpolation Module Model Architecture. Our interpolation module is built upon a U-Net architecture designed to generate temporal intermediate frames between adjacent frames in the video’s latent space.It takes as input the concatenated VAE- encoded latent features of two consecutive frames. The network consists of a three-level downsampling en- coder that progressively extracts high-level features, fol- lowed by a symmetric upsampling decoder that restores spa- tial resolution. Skip connections are employed to fuse high- resolution details from the encoder with deep semantic in- formation from the decoder. The final output is an interpo- lated latent feature map with the same number of channels as a single frame’s latent representation. Specifically,the encoder is composed of three ConvBlocklayers,eachprogressivelyexpanding the channel dimensions and followed by max-pooling for downsampling.The decoder utilizes transposed convo- lutions for upsampling, concatenates the feature maps from the corresponding encoder level, and then fuses the information using another ConvBlock. A final 1×1 con- volution is applied to produce the C-channel intermediate frame. Training Details. We train the U-Net interpolation mod- ule on a dataset of 150,000 real-world video clips, each 5 seconds long, with a resolution of 480p. The training pro- cedure is as follows: For each video, we first encode it into the latent space using the pre-trained VAE. A standard video in our latent space consists of 21 frames (corresponding to 81 frames at 16 fps in the pixel space over 5 seconds). From this 21- frame latent sequence, we uniformly downsample it to 11 frames. The U-Net module then takes these 11 frames and performs interpolation to restore the sequence to its original length. The network is optimized by computing a regression loss between the interpolated latent video and the original 21-frame ground-truth latent video. We use the AdamW optimizer with a learning rate of 1× 10 −5 , β 1 = 0.9, and β 2 = 0.999. The model is trained for 10,000 iterations with a batch size of 32. Inference Process. During inference, the interpolator is invoked along the temporal dimension. For each pair of ad- jacent frames in the input sequence, the original frames are preserved. The two frames are then concatenated and fed into the U-Net to generate one intermediate frame, which is inserted in order between them. This process expands the sequence length from F to 2F − 1, effectively doubling the video’s frame rate. The overall architecture is simple yet highly effective, capable of generating smooth and structurally consistent “Blue Lagoon, zoom out.“ “Burj Khalifa, zoom out.” t = 0s t = 1st =2s t =3s Figure 14. Failure cases resulting from unclipped temporal regularization. Without clipping, the student generator produces severe artifacts late in training. (Top) After the first second, a drastic content shift occurs, accompanied by a noticeable distortion of the building (highlighted by the red box). (Bottom) The scene content abruptly vanishes at the two-second mark and is replaced by another major content shift at the three-second mark (highlighted by the red box). These phenomena, inconsistent with plausible camera motion, are clear manifestations of hallucinations. This highlights the necessity of clipping the temporal loss to prevent it from excessively amplifying inter-frame variance. Frame Similarity Distribution Across Denoising Timesteps Cosine Similarity Denoising Timestep (Noise Level) Figure 15. Violin plot showing the distribution of adjacent-frame cosine similarity across different denoising stages. The overall similarity is notably higher during the high-noise stage, which mo- tivates our approach of halving the inference frame rate in this phase to reduce computational cost. temporal intermediates in the latent space. It is suitable for applications such as video frame-rate up-sampling, tran- sition generation, and interpolation tasks within diffusion models. 9. Hyperparameter Discuss In this section, we discuss the hyperparameter values used during the training process. For most settings, we follow the configuration from Wan2.1. The specific parameters are as follows: • Teacher Model: The number of training timesteps is set to 1000. The classifier-free guidance scale is 5.0, and the timestep-shift is also 5.0. • AdamW Optimizer: The parameters are set as follows: learning rate (lr) = 2.0×10 −6 , β 1 = 0.9, β 2 = 0.999, and the maximum gradient norm (maxgradnorm) is 10.0. • Two-Timescale Update Rule: The update frequency is set to 5. This means that the student generator is updated once for every five updates of the online model. 9.1. Temporal Regularization The temporal loss is updated using the formulation in Eq. (8); we simply set ε to 1×10 −6 . Theoretically, this loss function incentivizes a continuous increase in the temporal variance of the model’s output, as a larger variance results in a smaller negative logarithm and thus a lower loss. Con- sequently, the loss lacks a natural convergence mechanism. This can lead to numerical instability, excessively large gra- dients, and potentially exploding gradients. Furthermore, the model’s output can be amplified indefinitely, causing anomalous temporal variations that manifest as excessive jitter or noise in the generated videos. This regularization term may also conflict with the primary training objective; for instance, it creates an optimization paradox if the main task promotes smoothness while this term encourages vari- ance. Empirically, we observed that without any clipping, this loss causes the model to generate videos with severe frame jumps or hallucinatory artifacts in the later stages of training (as shown in Figure 14). Therefore, it is necessary to clip this loss once it falls below a certain value. To establish a reasonable clipping threshold, we first computed this loss on the VAE-encoded latents of 4,000 videos generated by the teacher model. The average value was found to be approximately 1.5. Based on this observation, we experimented with three distinct clip- ping thresholds: 1.2, 1.0, 0.6 and 0.4. We trained the model for a sufficient and equal number of epochs under each set- ting. The dynamic degree score and instance preservation score of the videos generated by the student are presented in Figure 12. We observed that setting the threshold to 0.6 significantly improves motion dynamics while maintaining object generation quality. Furthermore, to ensure that the distribution matching and adaptive regression losses remain the dominant sources of gradients during distillation, we found that setting the weight ω temp for this regularization term to 0.05 provides an effective balance.At this weight, the temporal loss typically converges to near the clipping threshold during training, while its weighted magnitude re- mains a small fraction compared to the other two losses. 9.2. Adaptive Regression Loss We now discuss the selection of key hyperparameters. First, the parameter α in Eq. (6) governs the contribu- tion of historical losses to the exponential moving average (EMA); we adopt a commonly used value of 0.95. Next, in Eq. (7), the parameter k controls the slope of the sigmoid function. This, in turn, determines how sensitively the adaptive weight responds to the deviation of the current lossL s from its historical trend ̄ L s . Notably, for low-noise steps, the absolute value of the regression loss L s is typi- cally very small (often below 0.01). At this stage, the guid- ance from the regression loss is inherently limited, as the input already contains significant information from the real image. Consequently, the output of the sigmoid function re- mains close to 0.5, making the weight largely insensitive to the value of k. Therefore, our analysis primarily focuses on the impact of k during the high-noise denoising stages. As visualized in Figure 13, which plots the weight func- tion for several values of k (where the x-axis is L s − ̄ L s ), we identified two desired behaviors. First, during the initial training phases, when the distribution gap is large,L s − ̄ L s can reach values between 0.6 and 0.8. We want to avoid excessively penalizing these data points (i.e., the weight should not be too close to 0 in this range). Second, in the later stages, when L s − ̄ L s has largely converged to the [-0.3, 0.3] interval, the weight function should still re- tain sufficient discriminative power to differentiate between samples. Through experimentation, we found that setting k = 3.0 provides an effective trade-off that satisfies both conditions. Finally, for w reg in Eq. (9), we set its value to 2.0. This choice normalizes the weight to approximately 1.0 for data points where the current loss is near the historical average (i.e.,L s − ̄ L s ≈ 0).