Pith. sign in

REVIEW 4 major objections 5 minor 78 references

Dual-Expert Consistency Model for Efficient and High-Quality Video Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Splitting consistency distillation into a semantic expert and a detail expert lets a 4-step video generator nearly match a 50-step teacher's quality.

desk verdict A genuinely useful dual-expert consistency distillation for video diffusion, with near-teacher quality at 4 steps, but the unverified boundary handoff and thin baseline set keep it from being fully convincing on its own. read the letter →

arxiv 2506.03123 v2 pith:QGANME4X submitted 2025-06-03 cs.CV

Zhengyao Lv , Chenyang Si , Tianlin Pan , Zhaoxi Chen , Kwan-Yee K. Wong , Yu Qiao , Ziwei Liu This is my paper
classification cs.CV
keywords videogenerationdiffusiondistillationconsistencymodelsfew-stepsamplingmodelaccelerationparameter-efficientfine-tuningtemporalcoherenceadversarialloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a single student model fails in few-step video distillation because consistency training across all noise levels pulls it in conflicting directions: high-noise steps demand semantic and layout decisions, while low-noise steps demand fine detail refinement. To resolve this, the authors propose the Dual-Expert Consistency Model, which splits the teacher's denoising trajectory at a boundary step and trains one expert for each half. A frozen semantic expert plus a small LoRA-based detail expert recovers nearly all of the teacher's quality at 4 steps: 83.83 versus 83.87 VBench on HunyuanVideo, and 79.99 versus 80.59 on CogVideoX, with the same per-step latency as prior few-step methods. The paper also introduces a temporal-coherence loss for the semantic expert and GAN plus feature-matching losses for the detail expert. If correct, it shows that decoupling distillation by noise phase is a viable route to cheap, high-quality video generation.

What carries the argument

The central object is the teacher's probability-flow ODE trajectory, the chain of noisy latents from pure noise to clean video. DCM splits this chain at a boundary $t_\kappa$ (default $N=50$, $\kappa=37$) into a semantic sub-trajectory and a detail sub-trajectory. The semantic expert is a full student trained on the high-noise segment with consistency loss plus a Temporal Coherence Loss that compares frame differences; the detail expert is initialized from it, frozen, and augmented with new timestep-dependent embedding layers and LoRA on the attention linear layers, trained on the low-noise segment with consistency, GAN, and Feature Matching losses. At inference the two experts are switched according to noise level, with steps evenly divided. This mechanism decouples the two learning tasks so the student is not forced to fit both high-gradient semantic changes and low-gradient detail refinements with one set of weights.

What would settle it

Swap the two training segments: train the detail expert on high-noise latents and the semantic expert on low-noise latents, then measure VBench. If the score does not fall well below the reported 83.83, the semantic/detail phase split is not the active ingredient in the gain.

Watch

Extended reading notes

Core claim

The paper claims that the optimization conflict in consistency distillation is what makes few-step video generation look degraded, and that the conflict can be removed by decoupling the student into two specialized experts. It presents evidence that a vanilla consistency model loses on both semantic and detail axes, while separate semantic and detail experts each beat it on their own segment and their combination is best. The parameter-efficient version preserves this with only timestep-dependent layers and LoRA added to the frozen semantic expert, and the expert-specific losses (Temporal Coherence, GAN, Feature Matching) each contribute measurable VBench gains.

Load-bearing premise

The load-bearing premise is that the teacher's denoising trajectory really does separate into a semantic phase and a detail phase at a single fixed boundary, with no need for the two experts to agree across that boundary.

Editorial extensions

If this is right

  • At 4 steps, DCM reaches 83.83 VBench on HunyuanVideo versus 83.87 for the 50-step teacher, and at 8 steps it reaches 83.86, so the acceleration is effectively lossless on this benchmark.
  • On CogVideoX-2B, 4-step DCM scores 79.99 versus the teacher's 80.59 and beats LCM (78.88) and PCM (79.09), showing the benefit transfers across model scales.
  • The parameter-efficient design adds only timestep-dependent layers and LoRA, so the second expert does not double memory or per-step compute at inference.
  • The expert-specific losses each contribute independently: Temporal Coherence improves motion consistency, and GAN plus Feature Matching improves detail realism.
  • DCM composes with sparse-attention acceleration (extra 1.33x speedup, VBench 83.79) and applies to a newer video model with minimal VBench loss, as reported in the supplementary material.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same semantic/detail decoupling could be tried for image diffusion distillation, where high-noise steps set global composition and low-noise steps add texture; the conflict may be smaller but structurally the same.
  • Because the detail expert is a small adapter on a frozen semantic expert, one could train several detail adapters for different style or quality targets and swap them at inference without retraining the base.
  • The fixed $\kappa=37$ boundary is picked from an L1-distance elbow on HunyuanVideo; choosing the boundary per prompt or per video, or enforcing cross-boundary consistency, could make the switch more robust and possibly push toward 2-step generation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper identifies a conflicting learning dynamics in consistency distillation for video diffusion models, where high-noise and low-noise steps demand different optimization behaviors. To address this, it proposes a Dual-Expert Consistency Model (DCM) that decouples the ODE trajectory into a semantic phase and a detail phase, training a semantic expert with consistency and temporal-coherence losses and a detail expert with consistency, GAN, and feature-matching losses. The detail expert is initialized from the frozen semantic expert with added time-dependent embedding layers and LoRA adapters. The method is evaluated on HunyuanVideo and CogVideoX, reporting 4-step VBench scores close to the 50-step teacher and higher than LCM and PCM baselines, with supplementary ablations and a user study supporting the design choices.

Significance. If the reported results are robust, DCM offers a practical and parameter-efficient way to accelerate large video diffusion models while preserving visual quality, which is a timely and industrially relevant contribution. The paper's strengths include a clear motivating analysis, a well-structured ablation study (Table 3) that isolates the decoupling, LoRA-based parameter efficiency, temporal coherence loss, and GAN loss, and a user study with a large number of raters. The release of code and models is a further positive factor. The central empirical claim, however, depends on a handoff between two independently trained experts whose boundary consistency is not verified, and on VBench point estimates that lack uncertainty quantification; these issues need to be addressed before the claim of matching the teacher can be fully accepted.

major comments (4)
  1. [Section 3.2, Eqs. (6)-(7) and Supplementary Section 7] The two experts are trained with separate consistency targets: SemE maps high-noise states to t_kappa and DetE maps low-noise states to t0. During inference, however, SemE runs first and hands its own boundary estimate \hat{x}_{t_kappa} to DetE, so DetE may receive inputs that are out-of-distribution relative to the teacher ODE boundary states it saw during training. The kappa ablation in Fig. 10 shows that the split point matters, but it does not measure the boundary mismatch or DetE's sensitivity to perturbed boundary states. Since the 4-step results in Table 1 depend on this handoff, the authors should provide evidence that SemE's boundary output is close to the teacher's, or add a consistency loss at the boundary, or report an experiment where DetE is applied to teacher boundary states versus SemE boundary states.
  2. [Table 1] All VBench scores are reported as single point estimates without variance, confidence intervals, or the number of seeds. The 4-step HunyuanVideo score (83.83) is within 0.04 of the 50-step teacher (83.87), and the CogVideoX 4-step score (79.99) is within 0.60 of its teacher (80.59); without multiple independent evaluations it is impossible to judge whether DCM truly matches the teacher or whether the difference is within sampling noise. Please report standard deviations or at least a deterministic evaluation protocol with fixed seeds, and clarify how many videos per prompt were used.
  3. [Related Work and Section 4.1] The paper cites Seaweed-APT [26] and the Motion Consistency Model [67] as prior video-distillation methods but does not compare against them in Table 1. If these baselines are not directly comparable because of different base models, resolutions, frame counts, or step regimes, that should be explicitly stated; as written, the abstract's claim of 'state-of-the-art visual quality' is not supported against the cited video-distillation literature. The authors should either add these comparisons where feasible or clearly delimit the claim to consistency-distillation methods on the two chosen backbones.
  4. [Section 4.1 and Supplementary Section 7] Several hyperparameters that directly affect the reported results are missing or chosen without ablation. Specifically, the lag l in the Temporal Coherence Loss (Eq. 8) is not defined; the LoRA rank and alpha are not reported; and the rule of 'evenly dividing the total steps' between experts in Supplementary Section 7 is stated as an empirical observation with no supporting ablation. Since the 4-step and 8-step numbers depend on these choices, the paper should report the exact values and, for the step-split rule, provide an ablation or a principled selection criterion.
minor comments (5)
  1. [Section 3.2 and Fig. 10] The procedure for selecting kappa=37 is described qualitatively ("decreases to a very small value"); please provide the exact threshold or a reproducible algorithm for choosing kappa from the L1 distance curve.
  2. [Eq. (8)] The temporal lag l in the Temporal Coherence Loss is introduced but never specified; please state its value in the implementation details.
  3. [Table 3] The column header 'GF' is not expanded in the caption; please define it as GAN + Feature Matching loss for clarity.
  4. [Section 4.1] For CogVideoX the training iterations are given as 'approximately 1000 steps' and 'around 500 steps'; please report exact numbers for reproducibility.
  5. [Supplementary Section 8.1] The acronym 'SVG' is used without expansion or citation at its first occurrence in the supplementary material; please define it and add the reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dual-expert design is empirically validated against external benchmarks and independent baselines; no derived result reduces to its own inputs.

full rationale

The paper's central claims are empirical and self-contained. The dual-expert consistency model is motivated by an analysis of teacher trajectory dynamics (Fig. 2) and loss/gradient differences, but the method itself is not derived from the evaluation metric. The consistency losses in Eqs. (6) and (7) are standard consistency-distillation objectives applied to sub-trajectories; they do not encode the reported VBench scores. The boundary kappa is selected from the teacher model's sampling dynamics (L1 distance between adjacent steps) rather than fitted to VBench, and the ablation in Fig. 10 is a hyperparameter study, not a prediction forced by construction. The temporal coherence, GAN, and feature-matching losses are added as regularizers and ablated individually; their contributions are measured, not assumed. Comparisons are made against independent baselines (LCM, PCM) and an external benchmark (VBench), and the authors also report a user study. Self-citations (e.g., VBench, FasterCache) appear in related work or as evaluation tools and are not load-bearing for the derivation. The potential train/inference mismatch at the SemE-to-DetE handoff is a robustness limitation, not a circular step: nothing in the losses or equations equates the final quality score to the training objective by definition. Therefore no circular step can be exhibited with the required specificity, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central method relies on standard consistency-distillation machinery and on an empirical two-phase decomposition of the teacher trajectory. The novel apparatus consists of a LoRA-based second expert and expert-specific losses; no new physical, mathematical, or external entities are introduced.

free parameters (5)
  • Trajectory boundary kappa = 37/50 for HunyuanVideo
    Chosen by inspecting the L1 distance between adjacent sampling steps (Section 3.2, Figure 10). It defines where the semantic expert ends and the detail expert begins, and VBench degrades when kappa deviates.
  • Inference step split between experts = 2/2 for 4 steps, 4/4 for 8 steps
    Supplementary Section 7 states that an even split was empirically observed to work well; it is not derived from the training boundary.
  • Temporal offset l in Temporal Coherence Loss = not reported
    Equation 8 uses a shift l along the temporal axis, but the value is never specified, so the motion-coherence tuning is not reproducible from the text.
  • LoRA rank and alpha for the detail expert = not reported
    The parameter-efficient design depends on these values, which are omitted from the implementation details.
  • Training iterations and learning rates for each expert = 1000 iters at 1e-6 and 1000 at 5e-6 for Hunyuan; about 1000/500 at 1e-6 for CogVideoX
    Manual hyperparameters chosen without a reported schedule search; they affect the final quality and are not swept.
assumptions (5)
  • domain assumption The teacher diffusion model's ODE solver trajectories are accurate enough to serve as distillation targets.
    All consistency losses in Equations 4, 6, and 7 are defined with respect to teacher predictions; no verification of solver error is provided.
  • ad hoc to paper Video generation separates cleanly into a semantic phase (layout and motion) at high noise and a detail phase at low noise.
    This two-phase structure is the paper's central hypothesis, introduced in Section 3.2 and used to justify the expert split; it is validated only by proxy, mainly the kappa sweep in Figure 10.
  • ad hoc to paper Independently distilled expert models that meet only at the boundary kappa produce a usable combined sampler.
    No consistency constraint is enforced between SemE and DetE at t_kappa; the even step split during inference is empirical (Supplementary Section 7).
  • domain assumption GAN and feature-matching losses using a frozen teacher backbone stabilize detail training without breaking consistency.
    Section 3.4 relies on this for training stability, but no convergence or mode-collapse analysis is presented.
  • standard math Standard definitions of the diffusion forward process, ODE solver, and consistency distillation are assumed from prior work.
    Equations 1 and 4 follow the standard diffusion and consistency-model formulations; the paper builds on them without re-deriving them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-Expert Consistency Model for Efficient and High-Quality Video Generation." pith.science (2026). https://pith.science/paper/QGANME4X

@misc{pith2026250603123,
  author       = {Pith},
  title        = {Pith review of: Dual-Expert Consistency Model for Efficient and High-Quality Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGANME4X}},
  note         = {Machine review of arXiv:2506.03123}
}
read the original abstract

Diffusion Models have achieved remarkable results in video synthesis but require iterative denoising steps, leading to substantial computational overhead. Consistency Models have made significant progress in accelerating diffusion models. However, directly applying them to video diffusion models often results in severe degradation of temporal consistency and appearance details. In this paper, by analyzing the training dynamics of Consistency Models, we identify a key conflicting learning dynamics during the distillation process: there is a significant discrepancy in the optimization gradients and loss contributions across different timesteps. This discrepancy prevents the distilled student model from achieving an optimal state, leading to compromised temporal consistency and degraded appearance details. To address this issue, we propose a parameter-efficient \textbf{Dual-Expert Consistency Model~(DCM)}, where a semantic expert focuses on learning semantic layout and motion, while a detail expert specializes in fine detail refinement. Furthermore, we introduce Temporal Coherence Loss to improve motion consistency for the semantic expert and apply GAN and Feature Matching Loss to enhance the synthesis quality of the detail expert.Our approach achieves state-of-the-art visual quality with significantly reduced sampling steps, demonstrating the effectiveness of expert specialization in video diffusion model distillation. Our code and models are available at \href{https://github.com/Vchitect/DCM}{https://github.com/Vchitect/DCM}.

Figures

Figures reproduced from arXiv: 2506.03123 by the authors.

Figure 1
Figure 1. Comparison of visual results between our DCM (4 steps), the original HunyuanVideo, and other competing methods (left). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the video synthesis process and the trend of loss variation. (a) In the early stages of sampling, the results change [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the visual quality of denoiser variants trained at different noise level samples. By optimizing two expert denoisers [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: The training process of DCM consists of two stages. In the semantic learning stage, we train SemE on high-noise samples with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Weight difference distribution between expert denoisers. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visual quality comparison of different methods. Differences are highlighted in boxes. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 10
Figure 10. Figure 10: Impact of different κ. based on the inference process [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 8
Figure 8. Figure 8: Impact of temporal coherence loss. Original w/o GAN & FM +FM +GAN +GAN+FM Local Zoom-in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Impact of the GAN loss and Feature Matching term. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 11
Figure 11. Figure 11: Visualization of the sampling process of different methods. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Visualization of the sampling process of different methods. [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Visual quality comparison of different methods. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Visual quality comparison of different methods. [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: Visual quality comparison of different methods. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Visual quality comparison of different methods. [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 18 canonical work pages

  1. [26]

    Diffusion adversarial post-training for one-step video generation.arXiv preprint arXiv:2501.08316,

    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,

    arXiv

  2. [67]

    Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation

    Yuanhao Zhai, Kevin Lin, Zhengyuan Yang, Linjie Li, Jian- feng Wang, Chung-Ching Lin, David Doermann, Junsong Yuan, and Lijuan Wang. Motion consistency model: Accel- erating video diffusion with disentangled motion-appearance distillation.arXiv preprint arXiv:2406.06890, 2024. 3

    work page Pith review arXiv 2024

  3. [1]

    Tract: Denoising diffusion models with transitive closure time-distillation.arXiv preprint arXiv:2303.04248, 2023

    David Berthelot, Arnaud Autef, Jierui Lin, Dian Ang Yap, Shuangfei Zhai, Siyuan Hu, Daniel Zheng, Walter Tal- bott, and Eric Gu. Tract: Denoising diffusion models with transitive closure time-distillation.arXiv preprint arXiv:2303.04248, 2023. 3

    arXiv 2023

  4. [2]

    Align your latents: High-resolution video synthesis with la- tent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dock- horn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with la- tent diffusion models. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 22563–22575, 2023. 2

    work page 2023

  5. [3]

    Nitrofusion: High-fidelity single-step diffu- sion through dynamic adversarial training.arXiv preprint arXiv:2412.02030, 2024

    Dar-Yen Chen, Hmrishav Bandyopadhyay, Kai Zou, and Yi-Zhe Song. Nitrofusion: High-fidelity single-step diffu- sion through dynamic adversarial training.arXiv preprint arXiv:2412.02030, 2024. 3

    arXiv 2024

  6. [4]

    Generative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian 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. 2, 6

    work page 2014

  7. [5]

    Boot: Data-free distillation of denoising diffusion models with bootstrapping

    Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Lingjie Liu, and Joshua M Susskind. Boot: Data-free distillation of denoising diffusion models with bootstrapping. InICML 2023 Work- shop on Structured Probabilistic Inference{\&}Generative Modeling, 2023. 3

    2023

  8. [6]

    Animatediff: Animate your personalized text- to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text- to-image diffusion models without specific tuning.arXiv preprint arXiv:2307.04725, 2023. 3

    arXiv 2023

Show all 78 references
  1. [7]

    Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weiss- buch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion....

    2024 arXiv

  2. [8]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2

    2020

  3. [9]

    Video dif- fusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif- fusion models.Advances in Neural Information Processing Systems, 35:8633–8646, 2022. 2

    2022

  4. [10]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 2

    2022 arXiv

  5. [11]

    Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021. 2, 5

    2021 arXiv

  6. [12]

    Animate anyone: Consistent and controllable image- to-video synthesis for character animation

    Li Hu. Animate anyone: Consistent and controllable image- to-video synthesis for character animation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024. 2

    2024

  7. [13]

    Dreamphysics: Learning physics-based 3d dynamics with video diffusion priors

    Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physics-based 3d dynamics with video diffusion priors. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3733–3741, 2025. 2

    2025

  8. [14]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pa...

    2024

  9. [15]

    Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling.arXiv preprint arXiv:2410.05954,

  10. [16]

    Distilling diffusion models into condi- tional gans

    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. InEuropean Conference on Computer Vision, pages 428–447. Springer, 2024. 3

    2024

  11. [17]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. 35:26565–26577, 2022. 2

    2022

  12. [18]

    Consistency trajectory mod- els: Learning probability flow ode trajectory of diffusion

    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. 3

    arXiv

  13. [19]

    Imagine flash: Accelerating emu diffusion models with backward distillation.arXiv preprint arXiv:2405.05224,

    Jonas Kohler, Albert Pumarola, Edgar Sch ¨onfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Tha- bet. Imagine flash: Accelerating emu diffusion models with backward distillation.arXiv preprint arXiv:2405.05224,

    arXiv

  14. [20]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 2, 4, 6, 7, 1

    2024 arXiv

  15. [21]

    Kling, 2024

    Kuaishou. Kling, 2024

    2024

  16. [22]

    Open-sora-plan, 2024

    PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2

    2024

  17. [23]

    Truncated con- sistency models.arXiv preprint arXiv:2410.14895, 2024

    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. 3

    2024 arXiv

  18. [24]

    Animatediff-lightning: Cross-model diffusion distillation.arXiv preprint arXiv:2403.12706, 2024

    Shanchuan Lin and Xiao Yang. Animatediff-lightning: Cross-model diffusion distillation.arXiv preprint arXiv:2403.12706, 2024. 3

    2024 arXiv

  19. [25]

    Sdxl- lightning: Progressive adversarial diffusion distillation

    Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxl- lightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3

    2024 arXiv

  20. [27]

    Scott: Accelerating diffusion models with stochastic consis- tency distillation.arXiv preprint arXiv:2403.01505, 2024

    Hongjian Liu, Qingsong Xie, Zhijie Deng, Chen Chen, Shix- iang Tang, Fueyang Fu, Zheng-jun Zha, and Haonan Lu. Scott: Accelerating diffusion models with stochastic consis- tency distillation.arXiv preprint arXiv:2403.01505, 2024. 3

    2024 arXiv

  21. [28]

    Pseudo numerical methods for diffusion models on manifolds.arXiv preprint arXiv:2202.09778, 2022

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds.arXiv preprint arXiv:2202.09778, 2022. 2

    2022 arXiv

  22. [29]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 3

    2022 arXiv

  23. [30]

    Instaflow: One step is enough for high-quality diffusion- based text-to-image generation

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion- based text-to-image generation. InThe Twelfth International Conference on Learning Representations, 2023. 3

    2023

  24. [31]

    Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024. 3

    2024 arXiv

  25. [32]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787,

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787,

  26. [33]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongx- uan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models.arXiv preprint arXiv:2211.01095, 2022. 2

    2022 arXiv

  27. [34]

    Knowledge distillation in iterative generative models for improved sampling speed

    Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 3

    2021 arXiv

  28. [35]

    Latent consistency models: Synthesizing high- resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    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. 3, 4, 6

    2023 arXiv

  29. [36]

    Lcm-lora: A universal stable-diffusion acceleration module.arXiv preprint arXiv:2311.05556, 2023

    Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolin´ario Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module.arXiv preprint arXiv:2311.05556, 2023. 3

    2023 arXiv

  30. [37]

    A comprehensive survey on knowledge distil- lation of diffusion models.arXiv preprint arXiv:2304.04262,

    Weijian Luo. A comprehensive survey on knowledge distil- lation of diffusion models.arXiv preprint arXiv:2304.04262,

    arXiv

  31. [38]

    Diff-instruct: A universal approach for transferring knowledge from pre-trained diffu- sion models.Advances in Neural Information Processing Systems, 36:76525–76546, 2023

    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. 3

    2023

  32. [39]

    One-step diffusion distillation through score implicit matching.arXiv preprint arXiv:2410.16794, 2024

    Weijian Luo, Zemin Huang, Zhengyang Geng, J Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching.arXiv preprint arXiv:2410.16794, 2024. 3

    2024 arXiv

  33. [40]

    You only sample once: Taming one-step text-to- image synthesis by self-cooperative diffusion gans.arXiv preprint arXiv:2403.12931, 2024

    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. 3

    2024 arXiv

  34. [41]

    Fastercache: Training-free video diffusion model acceleration with high quality.arXiv preprint arXiv:2410.19355, 2024

    Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality.arXiv preprint arXiv:2410.19355, 2024. 2

    2024 arXiv

  35. [42]

    Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024

    Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zi- wei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048, 2024. 2

    2024 arXiv

  36. [43]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 14297–14306, 2023. 3

    2023

  37. [44]

    Sora, 2024

    OpenAI. Sora, 2024. 2

    2024

  38. [45]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4195–4205,

  39. [46]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 3

    2022 arXiv

  40. [47]

    Hyper-sd: Trajectory segmented consistency model for efficient image synthesis

    Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024. 3, 4

    2024 arXiv

  41. [48]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1

    2022

  42. [49]

    Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022. 3

    2022 arXiv

  43. [50]

    Fast high- resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 3

    2024

  44. [51]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pages 87–103. Springer,

  45. [52]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1, 2

    2010 arXiv

  46. [53]

    Improved tech- niques for training consistency models.arXiv preprint arXiv:2310.14189, 2023

    Yang Song and Prafulla Dhariwal. Improved tech- niques for training consistency models.arXiv preprint arXiv:2310.14189, 2023. 3

    2023 arXiv

  47. [54]

    Consistency models.arXiv preprint arXiv:2303.01469, 2023

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models.arXiv preprint arXiv:2303.01469, 2023. 1, 3, 4

    2023 arXiv

  48. [55]

    Mochi 1.https :/ /github .com/ genmoai/models, 2024

    Genmo Team. Mochi 1.https :/ /github .com/ genmoai/models, 2024. 2

    2024

  49. [56]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    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. 1

    2025 arXiv

  50. [57]

    Phased consistency model.arXiv preprint arXiv:2405.18407, 2024

    Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingel- bach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency model.arXiv preprint arXiv:2405.18407, 2024. 3, 4, 6

    2024 arXiv

  51. [58]

    Animatelcm: Computation-efficient personalized style video generation without personalized video data

    Fu-Yun Wang, Zhaoyang Huang, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, and Hongsheng Li. Animatelcm: Computation-efficient personalized style video generation without personalized video data. InSIGGRAPH Asia 2024 Technical Communications, pages 1–5. 2024. 3

    2024

  52. [59]

    Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571, 2023. 3

    2023 arXiv

  53. [60]

    Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776, 2025

    Haocheng Xi, Shuo Yang, Yilong Zhao, Chenfeng Xu, Muyang Li, Xiuyu Li, Yujun Lin, Han Cai, Jintao Zhang, Dacheng Li, et al. Sparse videogen: Accelerating video diffusion transformers with spatial-temporal sparsity.arXiv preprint arXiv:2502.01776, 2025. 1

    2025 arXiv

  54. [61]

    Tack- ling the generative learning trilemma with denoising diffu- sion gans.arXiv preprint arXiv:2112.07804, 2021

    Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tack- ling the generative learning trilemma with denoising diffu- sion gans.arXiv preprint arXiv:2112.07804, 2021. 3

    2021 arXiv

  55. [62]

    Ufogen: You forward once large scale text-to-image gener- ation via diffusion gans

    Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image gener- ation via diffusion gans. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196–8206, 2024. 3

    2024

  56. [63]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024. 1, 2, 6

    2024 arXiv

  57. [64]

    Im- proved distribution matching distillation for fast image syn- thesis.arXiv preprint arXiv:2405.14867, 2024

    Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Im- proved distribution matching distillation for fast image syn- thesis.arXiv preprint arXiv:2405.14867, 2024. 3

    2024 arXiv

  58. [65]

    One-step diffusion with distribution matching distillation

    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

  59. [66]

    From slow bidirectional to fast causal video generators.arXiv preprint arXiv:2412.07772, 2024

    Tianwei Yin, Qiang Zhang, Richard Zhang, William T Free- man, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators.arXiv preprint arXiv:2412.07772, 2024. 3

    2024

  60. [68]

    Fast video generation with sliding tile attention.arXiv preprint arXiv:2502.04507, 2025

    Peiyuan Zhang, Yongqi Chen, Runlong Su, Hangliang Ding, Ion Stoica, Zhengzhong Liu, and Hao Zhang. Fast video generation with sliding tile attention.arXiv preprint arXiv:2502.04507, 2025. 2

    2025 arXiv

  61. [69]

    Faster video diffusion with trainable sparse attention.arXiv e-prints, pages arXiv–2505, 2025

    Peiyuan Zhang, Haofeng Huang, Yongqi Chen, Will Lin, Zhengzhong Liu, Ion Stoica, Eric P Xing, and Hao Zhang. Faster video diffusion with trainable sparse attention.arXiv e-prints, pages arXiv–2505, 2025. 2

    2025

  62. [70]

    Controlvideo: Training-free controllable text-to-video generation.arXiv preprint arXiv:2305.13077, 2023

    Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation.arXiv preprint arXiv:2305.13077, 2023. 2

    2023 arXiv

  63. [71]

    Videoelevator: El- evating video generation quality with versatile text-to-image diffusion models

    Yabo Zhang, Yuxiang Wei, Xianhui Lin, Zheng Hui, Peiran Ren, Xuansong Xie, and Wangmeng Zuo. Videoelevator: El- evating video generation quality with versatile text-to-image diffusion models. InProceedings of the AAAI Conference on Artificial Intelligence, pages 10266–10274, 2025

    2025

  64. [72]

    Framepainter: Endowing interactive image editing with video diffusion priors.arXiv preprint arXiv:2501.08225, 2025

    Yabo Zhang, Xinpeng Zhou, Yihan Zeng, Hang Xu, Hui Li, and Wangmeng Zuo. Framepainter: Endowing interactive image editing with video diffusion priors.arXiv preprint arXiv:2501.08225, 2025. 2

    2025 arXiv

  65. [73]

    Real-time video generation with pyramid attention broad- cast.arXiv preprint arXiv:2408.12588, 2024

    Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broad- cast.arXiv preprint arXiv:2408.12588, 2024. 2

    2024 arXiv

  66. [74]

    Fast sampling of dif- fusion models via operator learning

    Hongkai Zheng, Weili Nie, Arash Vahdat, Kamyar Aziz- zadenesheli, and Anima Anandkumar. Fast sampling of dif- fusion models via operator learning. InInternational con- ference on machine learning, pages 42390–42402. PMLR,

  67. [75]

    Trajectory consistency distillation.arXiv preprint arXiv:2402.19159, 2024

    Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation.arXiv preprint arXiv:2402.19159, 2024. 3

    2024 arXiv

  68. [76]

    Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation

    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. InForty-first International Confer- ence on Machine Learning, 2024. 3 Dual-Expert Co...

    2024

  69. [77]

    With 8 or 4 total steps, we assign 4 or 2 steps to each expert, re- spectively

    Further implementation details Stage division and expert switching.During inference, we empirically observe that evenly dividing the total steps between the two experts produces favorable results. With 8 or 4 total steps, we assign 4 or 2 steps to each expert, re- spectively. ...

  70. [78]

    Additional Results 8.1. Compatibility with other acceleration tech- niques DCM accelerates generation via sampling step reduction and is compatible with other methods like low precision computation and sparse modeling. For example, integrat- ing SVG [60] (which leverages the s...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.