pith. sign in

arxiv: 2606.18478 · v2 · pith:TDJ3TM3Znew · submitted 2026-06-16 · 💻 cs.CV

Data-Forcing Distillation: Restoring Diversity and Fidelity in Few-Step Video Generation

Siyi Chen , Shaowei Liu , Yixuan Jia , Zian Wang , Huan Ling , Qing Qu , Jun Gao This is my paper

Pith reviewed 2026-06-27 00:53 UTC · model grok-4.3

classification 💻 cs.CV
keywords data-forcing distillationfew-step video generationdiffusion model distillationmode collapseover-saturation artifactsteacher score discrepancyvideo synthesis
0
0 comments X

The pith

Data-Forcing Distillation restores diversity and fidelity in few-step video diffusion models via teacher score discrepancy guidance.

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

The paper targets two failure modes in prior distillation methods like DMD: reduced sample diversity from mode collapse and over-saturated outputs that stray from real video appearance. It introduces Data-Forcing Distillation as a minimal post-training adjustment that incorporates the teacher score discrepancy to steer the student model toward missing real-data modes while avoiding problematic saturated modes. This single-line modification requires only 100-300 finetuning steps yet improves both diversity and fidelity across text-to-video, image-to-video, and autoregressive settings on models such as Wan2.1-1.3B and Cosmos-Predict2.5-2B. The approach even yields outputs that surpass the original teacher in dynamics and appearance.

Core claim

DFD adds the teacher score discrepancy term to the reverse KL objective of distribution matching distillation, directing the student toward the real-data distribution to recover missing modes and suppress artifacts absent from real videos.

What carries the argument

Teacher score discrepancy term that measures the gap between teacher and student predictions to enforce real-data alignment.

If this is right

  • Few-step video generators regain sample diversity comparable to multi-step teachers.
  • Over-saturation artifacts are eliminated, yielding more natural video appearance and dynamics.
  • The method applies uniformly to text-to-video, image-to-video, and autoregressive video tasks.
  • Only 100-300 finetuning steps suffice to surpass the teacher model on tested architectures.
  • The change integrates into existing DMD-style pipelines with one line of code.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach may extend to distilling other generative modalities where reverse KL causes mode dropping.
  • Real-time video applications could benefit from the reduced step count while preserving fidelity.
  • Further analysis could test whether the discrepancy term scales to larger teacher-student gaps.

Load-bearing premise

The teacher score discrepancy term consistently guides the student to real data modes without introducing new artifacts or requiring architecture-specific adjustments.

What would settle it

Train a few-step student with DFD on a held-out video dataset and check whether mode collapse or over-saturation reappears in generated samples.

Figures

Figures reproduced from arXiv: 2606.18478 by Huan Ling, Jun Gao, Qing Qu, Shaowei Liu, Siyi Chen, Yixuan Jia, Zian Wang.

Figure 1
Figure 1. Figure 1: Overview of DFD. (a) DFD is a one-line change to DMD2, where the teacher scores a real video 𝑥 (w.p. 𝑝) instead of the student’s generation 𝐺𝜃(𝑧), giving the update 𝑔𝜃 ∝ 𝑠fake − 𝑠real. (b) This counters reverse-KL mode-seeking, recovering collapsed (diversity) and over-saturated off-manifold (fidelity) modes. (c) DFD improves quality and substantially increases diversity on text-to-video and image-to-video… view at source ↗
Figure 2
Figure 2. Figure 2: The comparison between our DFD and the origi￾nal DMD. Our DFD computes the real score directly using the videos sampled from the real data distribution, while the original DMD computes the real score using the gener￾ated videos from the student. The only difference compared to the DMD gradient (Eq. 3) is the second score: instead of evaluating ∇𝒙 log 𝑝real at the student’s own out￾put 𝐺𝜃(𝒛, 𝑐), we evaluate… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results on text-to-video generation. The left columns show models distilled on animation set, and the right column on the mix-style set. Our method produces videos that are not over-saturated, and recovers finer details such as the wing texture of the phoenix [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: I2V results on the ViPE test set. The colored outline indicates the input image. Our method produces videos that closely follow the first frame and remain coherent across the full sequence, whereas DMD2 and DP-DMD introduce structural anomalies such as an extra cowboy appearing from nowhere (highlighted by white circles). occasionally producing abrupt artifacts that disrupt the whole frame DMD2 yields visi… view at source ↗
Figure 5
Figure 5. Figure 5: I2V results on the VBench test set. The colored outline indicates the input image. Our method shows much better visual quality with smooth and physically plausible dynamics under complex scenarios, whereas DP-DMD and DMD2 fail to produce valid videos. 5.1.3 Autoregressive Video Generation Experimental Settings. For autoregressive video generation, we distill the Wan2.1-1.3B model on our curated mixed-style… view at source ↗
Figure 6
Figure 6. Figure 6: Autoregressive video generation results. Results from the generator trained with the Self Forcing pipeline using DFD loss outperform those trained with DMD loss, especially in frame consistency and physical plausibility [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a): Ablation on the effect of the GAN loss. Adding the GAN loss yields no clear quality improvement, and video dynamics even degrade, which is consistent with our quantitative results. (b): Ablation on the weight in Eq. 9: 𝑤 = 1 (upper) vs. 𝑤 = 1 2 (lower). Each row shows three evenly sampled frames from one generated video. There is no clear visual difference between the two settings. (c): Qualitative re… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of scaling up the distillation batch size. The colored outline indicates the input image. Increasing the batch size from 16 to 128 yields videos with superior visual quality and physical consistency. For instance, the spear in the rider’s hand exhibits much greater temporal clarity. Additionally, the scaled model preserves exact object counts across frames (the eggs), whereas the smaller batch size … view at source ↗
Figure 9
Figure 9. Figure 9: Limitation of DFD under a two-step generation budget. With two-step distillation of the Cosmos-Predict2.5-2B model, DFD still produces blurry results for fast-moving content (e.g., the man’s hands), loses fine detail (e.g., the woman’s face), or collapses to highly static videos. References [1] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based ge… view at source ↗
Figure 10
Figure 10. Figure 10: Diversity visualization across 8 random seeds for the same prompt. Columns 1–2: DMD2; columns 3–4: DP-DMD; columns 5–6: Ours. Each row shows the middle frame of two videos generated with different seeds. Our method produces visibly more diverse outputs across seeds. C.2 Main Results for Auto C.3 Additional Results for Image-to-Video Generation [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
read the original abstract

Recent progress has shown promise in distilling multi-step video diffusion models into efficient few-step students. Among them, Distribution Matching Distillation (DMD) and its successor DMD2 achieved strong generation quality and fast convergence. However, due to the nature of the reverse Kullback--Leibler (KL) objective, these methods exhibit two persistent failure modes: a substantial drop in sample diversity, and visibly over-saturated outputs that deviate from real-video appearance. In this work, we propose Data-Forcing Distillation (DFD), a simple post-training framework that restores diversity and fidelity in DMD with only a single-line of code change. At its core is the teacher score discrepancy to guide the student toward the real-data distribution, pulling it to missing modes (mitigating mode collapse) and away from problematic modes absent in real data (avoiding over-saturation). We provide an in-depth theoretical analysis of our framework and validate our approach on text-to-video, image-to-video, and autoregressive video generation. With only 100--300 steps of finetuning, DFD effectively restores diversity and fidelity on both Wan2.1-1.3B and Cosmos-Predict2.5-2B model, resolving the over-saturation artifacts with significantly better video dynamics and appearance, and even outperforms the teacher model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes Data-Forcing Distillation (DFD), a minimal post-training modification to Distribution Matching Distillation (DMD) for few-step video diffusion models. The core addition is a teacher score discrepancy term intended to restore sample diversity and correct over-saturation by guiding the student toward the real-data distribution. The manuscript provides theoretical analysis of the framework and reports empirical results on text-to-video, image-to-video, and autoregressive settings using Wan2.1-1.3B and Cosmos-Predict2.5-2B, claiming that 100-300 finetuning steps suffice to resolve the failure modes and that the resulting student can outperform the original multi-step teacher.

Significance. If the central claims hold under controlled evaluation, the work would be significant for practical few-step video generation: it targets a documented weakness of reverse-KL distillation methods with a lightweight fix and reports strong gains on two production-scale models. The emphasis on minimal additional training steps is a practical strength.

major comments (2)
  1. [Abstract, §5] Abstract and §5 (Experiments): The claim that DFD 'even outperforms the teacher model' is load-bearing for the central contribution. The reported comparisons must include a few-step teacher baseline evaluated at the same inference budget (4-8 steps) as the DFD student; without it, any apparent superiority is expected from the baseline multi-step vs. few-step mismatch and does not demonstrate that the score-discrepancy term itself yields a superior distribution.
  2. [§4] §4 (Theoretical analysis): The derivation that the teacher score discrepancy term 'pulls it to missing modes (mitigating mode collapse) and away from problematic modes' requires an explicit fixed-point or gradient analysis showing that the added term does not reduce to a reparameterization of the existing DMD objective or introduce new modes of collapse. The current description leaves open whether the guidance is parameter-free or requires model-specific scaling.
minor comments (2)
  1. [§5] Figure captions and §5 tables should explicitly state the number of denoising steps used for every baseline and proposed method to allow direct comparison of inference cost.
  2. [§3] The abstract states 'a single-line of code change'; the precise modified loss expression should appear in the main text (not only appendix) for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract, §5] Abstract and §5 (Experiments): The claim that DFD 'even outperforms the teacher model' is load-bearing for the central contribution. The reported comparisons must include a few-step teacher baseline evaluated at the same inference budget (4-8 steps) as the DFD student; without it, any apparent superiority is expected from the baseline multi-step vs. few-step mismatch and does not demonstrate that the score-discrepancy term itself yields a superior distribution.

    Authors: We agree that including a few-step teacher baseline at the same inference budget (4-8 steps) would strengthen the empirical validation and isolate the contribution of the score-discrepancy term. Our current comparisons focus on the original multi-step teacher to highlight recovery of its performance with few steps, but we acknowledge the referee's point regarding potential mismatch. In the revised version, we will add these controlled baselines for both Wan2.1-1.3B and Cosmos-Predict2.5-2B. revision: yes

  2. Referee: [§4] §4 (Theoretical analysis): The derivation that the teacher score discrepancy term 'pulls it to missing modes (mitigating mode collapse) and away from problematic modes' requires an explicit fixed-point or gradient analysis showing that the added term does not reduce to a reparameterization of the existing DMD objective or introduce new modes of collapse. The current description leaves open whether the guidance is parameter-free or requires model-specific scaling.

    Authors: We will expand §4 with an explicit fixed-point and gradient analysis of the combined DMD + teacher score discrepancy objective. This will show that the added term provides a distinct directional force toward the real-data distribution that is not equivalent to a reparameterization of the reverse-KL term alone. The core formulation is parameter-free, relying only on the difference in teacher scores; we will clarify this and any practical scaling factors used in experiments. revision: yes

Circularity Check

0 steps flagged

No circularity: new discrepancy term presented as independent addition with no reduction to fitted inputs or self-citations

full rationale

The paper introduces DFD as a post-training framework using a teacher score discrepancy term to restore diversity and fidelity in DMD-based distillation. The abstract and description frame this as a simple, independent modification (single-line code change) supported by in-depth theoretical analysis, without any equations or claims that reduce the new term to a fitted parameter from prior work, a self-citation chain, or a self-definitional loop. No load-bearing uniqueness theorems or ansatzes imported via citation are referenced in the provided text. The central claim of outperformance and restoration stands on the proposed term itself rather than re-labeling existing quantities. This is the common case of a self-contained method addition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No explicit free parameters, axioms, or invented entities are described in the abstract; the method relies on existing DMD setup plus the new discrepancy term.

pith-pipeline@v0.9.1-grok · 5786 in / 1144 out tokens · 27648 ms · 2026-06-27T00:53:02.982795+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 24 canonical work pages · 17 internal anchors

  1. [1]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  2. [2]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  3. [3]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn 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. 15

    2022

  4. [4]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    World Simulation with Video Foundation Models for Physical AI

    Arslan Ali, Junjie Bai, Maciej Bala, Yogesh Balaji, Aaron Blakeman, Tiffany Cai, Jiaxin Cao, Tianshi Cao, Elizabeth Cha, Yu-Wei Chao, and other. World simulation with video foundation models for physical AI.arXiv preprint arXiv:2511.00062, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Leo Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1(8):1, 2024

    2024

  9. [9]

    Progressive Distillation for Fast Sampling of Diffusion Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  10. [10]

    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 pattern recognition, pages 14297–14306, 2023

    2023

  11. [11]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023

    2023

  12. [12]

    Improved Techniques for Training Consistency Models

    YangSongandPrafullaDhariwal. Improvedtechniquesfortrainingconsistencymodels.arXiv preprint arXiv:2310.14189, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    Mean Flows for One-step Generative Modeling

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    AmirmojtabaSabour,SanjaFidler,andKarstenKreis.Alignyourflow: Scalingcontinuous-time flow map distillation.arXiv preprint arXiv:2506.14603, 2025

    work page arXiv 2025

  16. [16]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Free- man, and Taesung Park. One-step diffusion with distribution matching distillation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613–6623, 2024

    2024

  17. [17]

    Improved distribution matching distillation for fast image synthesis

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. Advances in neural information processing systems, 37:47455–47487, 2024

    2024

  18. [18]

    Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion 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 diffusion models.Advances in Neural Information Processing Systems, 36:76525–76546, 2023. 16

    2023

  19. [19]

    Uni-instruct: One-step diffusion model through unified diffusion divergence instruction.arXiv preprint arXiv:2505.20755, 2025

    Yifei Wang, Weimin Bai, Colin Zhang, Debing Zhang, Weijian Luo, and He Sun. Uni-instruct: One-step diffusion model through unified diffusion divergence instruction.arXiv preprint arXiv:2505.20755, 2025

    work page arXiv 2025

  20. [20]

    DreamFusion: Text-to-3D using 2D Diffusion

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  21. [21]

    Pro- lificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Pro- lificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in neural information processing systems, 36:8406–8441, 2023

    2023

  22. [22]

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

    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

    2014

  23. [23]

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback.Advances in neural information processing systems, 37:75692–75726, 2024

    Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Y Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback.Advances in neural information processing systems, 37:75692–75726, 2024

    2024

  25. [25]

    Consistency models made easy

    Zhengyang Geng, Ashwini Pokle, Weijian Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. InThe Thirteenth International Conference on Learning Representations, 2024

    2024

  26. [26]

    Consistency trajectory models: Learning probability flow ode trajectory of diffusion.arXiv preprint arXiv:2310.02279, 2023

    Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion.arXiv preprint arXiv:2310.02279, 2023

    work page arXiv 2023

  27. [27]

    Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency

    Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jintao 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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Phased consistency models.Advances in neural information processing systems, 37:83951–84009, 2024

    Fu-Yun Wang, Zhaoyang Huang, Alexander W Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency models.Advances in neural information processing systems, 37:83951–84009, 2024

    2024

  29. [29]

    arXiv preprint arXiv:2502.15681 , year=

    Yilun Xu, Weili Nie, and Arash Vahdat. One-step diffusion models with𝑓-divergence distri- bution matching.arXiv preprint arXiv:2502.15681, 2025

    work page arXiv 2025

  30. [30]

    One-step diffusiondistillationthroughscoreimplicitmatching.AdvancesinNeuralInformationProcessing Systems, 37:115377–115408, 2024

    Weijian Luo, Zemin Huang, Zhengyang Geng, J Zico Kolter, and Guo-jun Qi. One-step diffusiondistillationthroughscoreimplicitmatching.AdvancesinNeuralInformationProcessing Systems, 37:115377–115408, 2024

    2024

  31. [31]

    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 Conference on Machine Learning, 2024

    2024

  32. [32]

    Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis

    Tianhe Wu, Ruibin Li, Lei Zhang, and Kede Ma. Diversity-preserved distribution matching distillation for fast visual synthesis.arXiv preprint arXiv:2602.03139, 2026. 17

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    Transition matching distillation for fast video generation.ArXiv, abs/2601.09881, 2026

    Weili Nie, Julius Berner, Nanye Ma, Chao Liu, Saining Xie, and Arash Vahdat. Transition matching distillation for fast video generation.ArXiv, abs/2601.09881, 2026

    work page arXiv 2026

  34. [34]

    Training neural samplers with reverse diffusive kl divergence.arXiv preprint arXiv:2410.12456, 2024

    Jiajun He, Wenlin Chen, Mingtian Zhang, David Barber, and José Miguel Hernández-Lobato. Training neural samplers with reverse diffusive kl divergence.arXiv preprint arXiv:2410.12456, 2024

    work page arXiv 2024

  35. [35]

    Adversarial distribution matching for diffusion distillation towards efficient image and video synthesis

    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 distillation towards efficient image and video synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16818–16829, 2025

    2025

  36. [36]

    ViPE: Video Pose Engine for 3D Geometric Perception

    Jiahui Huang, Qunjie Zhou, Hesam Rabeti, Aleksandr Korovko, Huan Ling, Xuanchi Ren, Tianchang Shen, Jun Gao, Dmitry Slepichev, Chen-Hsuan Lin, Jiawei Ren, Kevin Xie, Joydeep Biswas,LauraLeal-Taixe,andSanjaFidler.Vipe: Videoposeenginefor3dgeometricperception. InNVIDIA Research Whitepapers arXiv:2508.10934, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    Vbench: Comprehensive benchmark suiteforvideogenerativemodels

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suiteforvideogenerativemodels. InProceedingsoftheIEEE/CVFConferenceonComputerVision and Pattern Recognition, pages 21807–21818, 2024

    2024

  38. [38]

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  40. [40]

    0,2,5,7,9,12,...,38,39

    AlecRadford,JongWookKim,ChrisHallacy,AdityaRamesh,GabrielGoh,SandhiniAgarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual modelsfromnaturallanguagesupervision. InInternationalconferenceonmachinelearning,pages 8748–8763. PmLR, 2021. 18 A Additional Theory A.1 teacher score discrepancy A good regularizer mus...

    work page arXiv 2021