Pith. sign in

REVIEW 3 major objections 6 minor 86 references

A 5-billion-parameter video diffusion transformer can run on a phone and generate 5-second clips in 20 seconds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 08:16 UTC pith:2F4ETHR7

load-bearing objection Real 5B-on-phone engineering, but the 'quality gap' claim is undercut by the paper's own user study and a mismatch between the evaluated and final model. the 3 major comments →

arxiv 2607.06173 v2 pith:2F4ETHR7 submitted 2026-07-07 cs.CV

MobileWan: Closing the Quality Gap for Mobile Video Diffusion

Mohsen Ghafoorian , Denis Korzhenkov , Adil Karjauv , Ioannis Lelekas , Noor Fathima , Spyridon Stasis , Hanno Ackermann , Boris van Breugel
show 4 more authors
Markus Nagel Fatih Porikli Animesh Karnewar Amirhossein Habibian
This is my paper
classification cs.CV
keywords video diffusionmobile NPU inferencerecurrent attentionlinear attentionattention head pruningstep distillationVAE decoder optimizationconstant-memory generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to close the quality gap between mobile and server video generation by showing that a 5-billion-parameter diffusion transformer — the scale normally reserved for server GPUs — can run on a phone. It does so by converting attention into a chunk-wise recurrent process with constant memory, pruning attention heads with learned binary gates, distilling sampling to three steps, and replacing the heavy VAE decoder with a lightweight one. The result is a system that generates 5-second 480x832 clips at 16 FPS in 20 seconds end-to-end on a mobile NPU, with a VBench score of 83.79, competitive with the server baseline. If correct, this reframes mobile video generation: quality is no longer bounded by the 0.4–1.8B parameter budgets of prior mobile models.

Core claim

MobileWan claims a server-scale 5-billion-parameter video diffusion transformer can run on a phone without sacrificing quality. All blocks are distilled into hybrid local-softmax/causal-linear attention, so inference becomes an RNN with two fixed-size accumulators carrying all past-chunk context; memory no longer grows with video length. Learned binary per-head gates, a three-step sampler, and a lightweight decoder with extended look-back complete the pipeline. The system generates 5-second 480x832 clips at 16 FPS in 20 seconds on a mobile NPU, reaching VBench 83.79 versus the server baseline's 83.12 — the claimed first 5B-scale mobile deployment.

What carries the argument

The load-bearing object is a recurrent hybrid attention layer: within each temporal chunk, tokens attend via softmax to a small local window, and all earlier chunks are summarized by two fixed-size accumulators (s_t, z_t) updated causally through linear-attention feature maps. Because the linear path is strictly causal, the trained transformer can be executed as a stateful RNN with constant memory. The second mechanism is binary per-head gating: learnable scalar gates, annealed and then frozen, decide which self-attention heads survive, with training biased toward high-noise timesteps to make pruning decisions robust.

Load-bearing premise

The deployment claim rests on the assumption that the research-prototype phone used for measurements, with about 9.9 GB peak RAM for one sample, represents the memory and thermal budget of a shipping commercial device; if a real phone offers less NPU memory or throttles under sustained load, the 20-second end-to-end result will not transfer.

What would settle it

Run the released checkpoint on a retail phone, not a prototype, under repeated generation until thermal equilibrium and measure peak RAM and end-to-end latency; if peak RAM exceeds available NPU memory or latency climbs well above 20 seconds after warm-up, the central deployment claim is refuted.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Mobile video generation no longer needs sub-2B backbones: the 5B model's VBench total (83.79) sits between the server baseline (83.12) and the prior mobile state of the art (81.61).
  • Because the recurrent state has fixed size, memory stays roughly constant as video length grows, so longer clips should cost mainly time, not RAM.
  • Full recurrence cuts self-attention memory from 15.7 GB to 0.1 GB, allowing the whole 30-block transformer to load once rather than in splits, removing split-initialization overhead.
  • A blinded user study prefers the system over the prior mobile pipeline 80% of the time and rates it no worse than the server teacher in a quarter of paired comparisons.

Where Pith is reading between the lines

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

  • We infer the same recurrent-distillation recipe could generalize to other pretrained generative transformers, potentially closing similar quality gaps for image and audio generation on-device.
  • The success of high-noise-biased gate training suggests a transferable principle: pruning decisions should be made under timesteps that emphasize global structure, which could improve aggressive compression in other diffusion models.
  • The admitted temporal discontinuities hint that the fixed-size causal state is lossy; a natural test is to lengthen chunk overlap or add learned state compression and measure whether long-range coherence improves.
  • Retail-phone thermal testing is the obvious next experiment: sustained generation on a shipping device, not a prototype, will determine whether the 20-second latency is a true product feature or a lab artifact.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces MobileWan, a system that adapts Wan2.2-5B, a server-scale video diffusion transformer, for deployment on mobile NPUs. The technical recipe combines: (i) a noise-biased learnable attention-head pruning method; (ii) recurrence distillation that converts the transformer into a chunk-wise autoregressive RNN using causal linear attention with fixed-size state; (iii) step distillation (DMD/D-DMD/DMD2 variants); and (iv) a memory-optimized causal decoder. The authors report 5-second 480x832 videos at 16 FPS in ~20 seconds on a Snapdragon 8 Gen. 5 NPU, with a VBench total score of 83.79, which they claim closes the mobile-vs-server quality gap. The paper includes system measurements, a user study, and comparisons against prior mobile and server video diffusion models.

Significance. If the technical claims hold, this is a significant engineering and algorithmic contribution: it demonstrates that a 5B-parameter video diffusion transformer can be made to run on mobile-class NPU hardware via recurrent reformulation, without resorting to small models. The core RNN reformulation is mathematically sound as a linear-attention identity, and the system-level measurements (latency, memory, sparsity) are detailed and useful. The release of the checkpoint and code is a concrete asset. However, the headline claim of 'closing the quality gap' is not supported by the paper's own human-preference evidence, and the 'commercial mobile device' claim rests on measurements from a research prototype. The significance is therefore real but the central quality claim requires substantial revision or additional evidence.

major comments (3)
  1. [§4, Table 6; §3.3] The human-preference study does not support the 'closing the quality gap' claim. Table 6 shows that Wan2.2 5B-FT is preferred 53% to 22% over MobileWan (25% no preference), and the text's statement that the system 'in half of the cases is not performing worse' is misleading (22% + 25% = 47%, and 'no preference' is not a preference for MobileWan). Moreover, §3.3 states the user study used the D-DMD model with 3 sampling steps (VBench 82.44 in Table 3), while the final integrated system in Table 5 uses the DMD2-based step distillation (VBench 84.03/83.79). The reported human judgments therefore do not evaluate the final MobileWan model whose quality is claimed. Combined with the Appendix A.6 admission that VBench rewards oversaturated colors and may not align with human preference, the central quality claim rests on an automated metric that the paper itself cautions against.
  2. [Appendix A.7; Abstract] The abstract and introduction claim deployment on 'a commercial mobile device,' but the only on-device measurements are taken on a 'research prototype device with a Qualcomm Snapdragon 8 Gen. 5 Mobile Platform' (A.7). Table 11 reports a peak RAM of 9.9 GB for a single sample with the full-recurrence DiT, and no sustained-thermal, power, or battery measurements are provided. As written, the evidence does not establish deployment on a shipping commercial phone; it establishes feasibility on a prototype evaluation platform. The claim should be qualified accordingly or substantiated with measurements on a production device.
  3. [§5; §3.2] The abstract claims the method 'preserves temporal coherence across chunks,' but the Conclusion explicitly states that 'in certain cases our RNN reformulation leads to temporal discontinuities in the generated scenes' and that the step-distillation/decoder combination causes flickering. These are not peripheral: they directly affect the core benefit of the recurrent formulation. The paper should either quantify the frequency/severity of these discontinuities or weaken the abstract and Section 3.2 claims to 'partially preserves' / 'with occasional discontinuities.' Without this, the presentation overstates the demonstrated behavior of the central mechanism.
minor comments (6)
  1. [Table 5] Typo: 'Reccurence Distil.' should be 'Recurrence Distil.'.
  2. [§4, Human Evaluation] The phrasing 'in half of the cases is not performing worse' should be replaced with the actual numbers from Table 6 (22% preferred MobileWan, 25% no preference, 53% preferred baseline).
  3. [§3.3] The relationship between the user-study D-DMD model and the final DMD2-based system should be clarified in the main text, not only implied by the table values. Readers should be told explicitly that the final model's human preferences were not measured.
  4. [Appendix A.6] The caveat that 'VBench toolkit gives higher scores to slightly oversaturated colors, and this does not always align with human preference' should appear in the main text wherever VBench is used as the headline quality metric.
  5. [Table 7] The S2DiT - AR model reports VBench 83.26, close to MobileWan's 83.79. The claim of 'new state of the art' should acknowledge that the margin is small and may be within benchmark noise, given the VBench caveats.
  6. [References] Reference [40] contains a placeholder year '202X'; please fill in the correct year.

Circularity Check

0 steps flagged

No significant circularity: the central derivation is self-contained and externally benchmarked.

full rationale

The paper's claimed chain — Wan2.2-5B -> head pruning -> recurrence distillation -> step distillation -> decoder optimization -> on-device deployment — is not circular by construction. Each stage has a stated objective (flow-matching loss, sparsity regularizer, teacher-student distillation, reconstruction + perceptual loss) and the final claims are checked against external measurements: VBench is an independent benchmark, and the latency/RAM numbers are hardware measurements on a Snapdragon 8 Gen 5 NPU. The reported VBench 83.79 is a measured value, not a term defined by the method's own equations or fitted parameters. The self-citations to the same group's prior work ([15], [16], [1]) supply building blocks such as the hybrid linear/softmax attention recipe and polynomial feature maps, but the paper also specifies the architecture and validates it independently; no uniqueness theorem is imported to rule out alternatives, and the basic RNN view of causal linear attention is attributed to external work [57]. The manuscript does contain evidence-quality limitations that weaken the strong 'closing the quality gap' and 'commercial mobile device' claims: A.6 concedes VBench rewards oversaturated colors and can disagree with human preference; A.7 states measurements were made on research prototype devices; the human study used the D-DMD model rather than the final DMD2 variant, and Table 6 shows the Wan2.2 baseline was preferred 53% vs 22%; Section 5 admits temporal discontinuities, flickering, and reduced motion. These are correctness/evidence concerns, not circularity: none of them amounts to a prediction being equivalent to an input by definition. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on a handful of hand-chosen architectural hyperparameters (pruning ratio, chunk size, overlap, noise-bias distributions, quantization bits) and on the empirical assumption that distillation on synthetic Wan2.1-generated data transfers to real prompt distributions. No new physical entities or unexplained theoretical objects are introduced.

free parameters (6)
  • Head pruning ratio = 23%
    Chosen from Figure 3 as the operating point balancing FLOPs reduction and VBench; not derived.
  • Recurrence chunk size T_s = 4 temporal slices
    Architectural hyperparameter set in A.3.2; controls the local softmax window and the RNN state update size.
  • Chunk overlap T_o = 2 temporal slices
    Hand-selected in A.3.2 to improve temporal coherence; interacts with the causality boundary.
  • Noise-bias LogitNormal parameters for gate learning = LogitNormal(1.5, 1) for phases 1-2; LogitNormal(0, 1) after binarization
    Selected based on Table 1; high-noise-biased training was found critical for gate learning.
  • DMD/D-DMD noise-level prior = Shifted uniform distribution with parameter 5
    Chosen for step-distillation sampling; affects the trade-off between color fidelity and motion.
  • Quantization bit widths = 8-bit weights, mixed-precision activations (16-bit for residuals)
    Engineering choice described in Section 4; required for the reported on-device RAM/latency.
axioms (4)
  • standard math The exponential softmax kernel can be approximated by finite polynomial feature maps with negligible drift over long video generation.
    Inherited from ReHyAt [15,16]; if the kernel approximation error compounds across chunks, the RNN outputs would drift, but no error bounds are provided.
  • domain assumption A pretrained softmax DiT can be block-wise distilled into a recurrent hybrid form without catastrophic quality loss.
    Validated empirically with VBench, but not guaranteed; the paper's own radar plot notes temporal jitter in the full-RNN variant.
  • domain assumption Fine-tuning on 80K synthetic videos generated by Wan2.1 14B is a sufficient training distribution for the target resolution and quality.
    The entire fine-tuning and decoder training uses this in-house synthetic dataset (A.4, Table C); no external validation set beyond VBench/DAVIS is used.
  • domain assumption Measurements on a Snapdragon 8 Gen 5 research prototype are representative of a commercial mobile device.
    Appendix A.7 states the results are from 'research prototype devices'; the abstract and introduction claim deployment on a 'commercial mobile device' without validating thermal or shipping configurations.

pith-pipeline@v1.3.0-alltime-deepseek · 21267 in / 10795 out tokens · 102296 ms · 2026-08-02T08:16:03.241093+00:00 · methodology

0 comments
read the original abstract

Recent advances in video diffusion have been driven by scaling transformer-based architectures to billions of parameters, substantially improving visual fidelity and motion coherence. In contrast, existing mobile video diffusion models remain limited to relatively small parameter budgets, typically 0.4-1.8B, restricting generation quality. In this work, we show that high-quality mobile video generation does not require small models. Instead, we demonstrate that a server-scale 5B-parameter video diffusion transformer can be deployed efficiently on memory-constrained mobile hardware through recurrent reformulation and structured compression. Starting from Wan2.2-5B, we rely on a recurrence distillation framework that converts video generation into a chunk-wise autoregressive process with constant-memory attention computation. Combined with causal linear attention, the model operates as an RNN at inference time while preserving temporal coherence across chunks. We further propose a learnable attention head pruning method based on binary per-head gates optimized end-to-end using a noise-biased sparsity objective and distillation-based finetuning. Together with sampling-step distillation and memory-optimized VAE decoding, MobileWan becomes the first 5B-scale video diffusion model deployable on a commercial mobile device. Our system generates 5-second 480x832 videos at 16 FPS in 20 seconds end-to-end latency, achieving a VBench score of 83.79 and establishing a new state of the art in mobile video generation. Please find the released DiT checkpoint and the sampling code in the project page: https://qualcomm-ai-research.github.io/MobileWan

Figures

Figures reproduced from arXiv: 2607.06173 by Adil Karjauv, Amirhossein Habibian, Animesh Karnewar, Boris van Breugel, Denis Korzhenkov, Fatih Porikli, Hanno Ackermann, Ioannis Lelekas, Markus Nagel, Mohsen Ghafoorian, Noor Fathima, Spyridon Stasis.

Figure 1
Figure 1. Figure 1: We significantly advance the quality of Mobile Video Generation compared to the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MobileWan: We apply a sequence of complementary optimizations, targeting both the diffusion transformer and the decoder, to the Base Wan 2.2 5B model which requires a typical Server GPU like 80GB-A100 and transform it into MobileWan running natively on a mobile device powered by Snapdragon® 8 Gen. 5 NPU. 1 Introduction Video diffusion models have rapidly become a central paradigm for generative visual mode… view at source ↗
Figure 3
Figure 3. Figure 3: VBench comparison of head pruning meth￾ods. Learnable pruning with high-noise-biased train￾ing performs best, especially at aggressive pruning ratios [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the VAE architecture. We retain the Wan2.2 encoder and modify the MemBlock from LightX2V [35] by extending the causal look-back window at each layer, enabling the prediction of frame 𝐹𝑡 to leverage information from up to 𝐹𝑡−4 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Evaluation of MobileWan model. We visualize uniformly selected frames from the 81 × 480 × 832 videos generated on Qualcomm Snapdragon® 8 Gen. 5 NPU. tokens, resulting in a drastic reduction in attention weights. This enables a denoising step over the entire video to be completed in 6.6 seconds. Human Evaluation We extend our quantitative evaluation with two internal user studies. In the first s… view at source ↗
Figure 6
Figure 6. Figure 6: Head Pruning FLOPs vs Pruning Ratio. Number of FLOPs and the corresponding reduction at different pruning ratios with respect to the baseline [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Radar plot comparing a subset of our hybrid models with the original Wan2.2 5B model on the full [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Our optimized pipeline matches exactly the base Wan 2.2 model’s instantiation of the latent video [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt: A bear and a zebra [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompt: A dog running happily [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prompt: A happy fuzzy panda playing guitar nearby a campfire, snow mountain in the background [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prompt: A person eating a burger [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Prompt: A person is pushing up [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prompt: A teddy bear is swimming in the ocean [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Full set of VBench dimensions for optimizations steps used by [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗

Review history (2 revisions) →

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

86 extracted references · 11 linked inside Pith

  1. [1]

    InICLR, 2026

    Animesh Karnewar, Denis Korzhenkov, Ioannis Lelekas, Noor Fathima, Adil Karjauv, Mohsen Ghafoorian,andAmirhosseinHabibian.Neodragon: Mobilevideogenerationusingdiffusiontransformer. InICLR, 2026

    2026

  2. [2]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/index/ video-generation-models-as-world-simulators/, 2024. OpenAI technical report, accessed 2026-05-05

    2024

  3. [3]

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

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

    Pith/arXiv arXiv 2024

  4. [4]

    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, Da Yin, Zhang Yuxuan, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. InICLR, 2025

    2025

  5. [5]

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

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

    Pith/arXiv arXiv 2025

  6. [6]

    Efficient video diffusion models: Advancements and challenges.arXiv preprint arXiv:2604.15911, 2026

    Shitong Shao, Lichen Bai, Pengfei Wan, James Kwok, and Zeke Xie. Efficient video diffusion models: Advancements and challenges.arXiv preprint arXiv:2604.15911, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    Moalign: Motion-centric representation alignment for video diffusion models.arXiv preprint arXiv:2510.19022, 2025

    Aritra Bhowmik, Denis Korzhenkov, Cees GM Snoek, Amirhossein Habibian, and Mohsen Ghafoo- rian. Moalign: Motion-centric representation alignment for video diffusion models.arXiv preprint arXiv:2510.19022, 2025

    arXiv 2025

  8. [8]

    A systematic post-train framework for video generation.arXiv preprint arXiv:2604.25427, 2026

    ZeyueXue,SimingFu,JieHuang,ShuaiLu,HaoranLi,YijunLiu,YumingLi,XiaoxuanHe,Mengzhao Chen, Haoyang Huang, et al. A systematic post-train framework for video generation.arXiv preprint arXiv:2604.25427, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Mobile-oriented video diffusion: Enabling text-to-video generation on mobile devices without retraining, compression, or pruning

    Bosung Kim, Kyuhwan Lee, Isu Jeong, Jungmin Cheon, Yeojin Lee, and Seulki Lee. Mobile-oriented video diffusion: Enabling text-to-video generation on mobile devices without retraining, compression, or pruning. InWACV, 2026

    2026

  10. [10]

    Tamingdiffusiontransformer for real-time mobile video generation.arXiv preprint arXiv:2507.13343, 2025

    Yushu Wu, Yanyu Li, Anil Kag, Ivan Skorokhodov, Willi Menapace, Ke Ma, Arpit Sahni, Ju Hu, AliaksandrSiarohin,DhritimanSagar,YanzhiWang,andSergeyTulyakov. Tamingdiffusiontransformer for real-time mobile video generation.arXiv preprint arXiv:2507.13343, 2025

    arXiv 2025

  11. [11]

    S2dit: Sandwichdiffusiontransformerformobilestreaming video generation.arXiv preprint arXiv:2601.12719, 2026

    Lin Zhao, Yushu Wu, Aleksei Lebedev, Dishani Lahiri, Meng Dong, Arpit Sahni, Michael Vasilkovsky, HaoChen,JuHu,AliaksandrSiarohin,etal. S2dit: Sandwichdiffusiontransformerformobilestreaming video generation.arXiv preprint arXiv:2601.12719, 2026

    arXiv 2026

  12. [12]

    Mobile video diffusion

    Haitam Ben Yahia, Denis Korzhenkov, Ioannis Lelekas, Amir Ghodrati, and Amirhossein Habibian. Mobile video diffusion. InICCV, 2025

    2025

  13. [13]

    Snapgen-v: Generating a five-second video within five seconds on a mobile device

    YushuWu, ZhixingZhang, YanyuLi, YanwuXu, AnilKag, YangSui, HuseyinCoskun, KeMa, Aleksei Lebedev, Ju Hu, et al. Snapgen-v: Generating a five-second video within five seconds on a mobile device. InCVPR, 2025

    2025

  14. [14]

    Sana-video: Efficient video generation with block linear diffusion transformer

    Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, et al. Sana-video: Efficient video generation with block linear diffusion transformer. InProceedings of the International Conference on Learning Representations (ICLR), 2026

    2026

  15. [15]

    Attention surgery: An efficient recipe to linearize your video diffusion transformer

    Mohsen Ghafoorian, Denis Korzhenkov, and Amirhossein Habibian. Attention surgery: An efficient recipe to linearize your video diffusion transformer. InCVPR, 2026

    2026

  16. [16]

    Rehyat: Recurrent hybrid attention for video diffusion transformers

    Mohsen Ghafoorian and Amirhossein Habibian. Rehyat: Recurrent hybrid attention for video diffusion transformers. InCVPR, 2026

    2026

  17. [17]

    Movie: Mobile diffusion for video editing.arXiv preprint arXiv:2412.06578, 2024

    Adil Karjauv, Noor Fathima, Ioannis Lelekas, Fatih Porikli, Amir Ghodrati, and Amirhossein Habibian. Movie: Mobile diffusion for video editing.arXiv preprint arXiv:2412.06578, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Clockwork diffusion: Efficient generation with model-step distillation

    Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In CVPR, 2024

    2024

  19. [19]

    Amd-hummingbird: Towards an efficient text-to-video model.arXiv preprint arXiv:2503.18559, 2025

    Takashi Isobe, He Cui, Dong Zhou, Mengmeng Ge, Dong Li, and Emad Barsoum. Amd-hummingbird: Towards an efficient text-to-video model.arXiv preprint arXiv:2503.18559, 2025

    arXiv 2025

  20. [20]

    Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    2023

  21. [21]

    On-device sora: Enabling training-free diffusion-based text-to-video generation for mobile devices.arXiv preprint arXiv:2502.04363, 2025

    Bosung Kim, Kyuhwan Lee, Isu Jeong, Jungmin Cheon, Yeojin Lee, and Seulki Lee. On-device sora: Enabling training-free diffusion-based text-to-video generation for mobile devices.arXiv preprint arXiv:2502.04363, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    SANA 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng YU, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, Bingchen Liu, Daquan Zhou, and Song Han. SANA 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. InICML, 2025

    2025

  23. [23]

    FastLightGen: Fast and light video generation with fewer steps and parameters

    Shitong Shao, Yufei Gu, and Zeke Xie. FastLightGen: Fast and light video generation with fewer steps and parameters. InCVPR, 2026

    2026

  24. [24]

    Hadzic, Manling Li, AgrimGupta,StefanoMassaroli,AzaliaMirhoseini,JuanCarlosNiebles,StefanoErmon,andLiFei-Fei

    Keshigeyan Chandrasegaran, Michael Poli, Daniel Y Fu, Dongjun Kim, Lea M. Hadzic, Manling Li, AgrimGupta,StefanoMassaroli,AzaliaMirhoseini,JuanCarlosNiebles,StefanoErmon,andLiFei-Fei. Exploring diffusion transformer designs via grafting. InNeurIPS, 2026

    2026

  25. [25]

    Freeman, and Taesung Park

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step Diffusion with Distribution Matching Distillation. InCVPR, 2024

    2024

  26. [26]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  27. [27]

    Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield

    Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Steven HOI, and Hongsheng Li. Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield. InICLR, 2026

    2026

  28. [28]

    Barron, and Ben Mildenhall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. InICLR, 2023

    2023

  29. [29]

    Jacobs, Alexei A

    David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking Score Distillation as a Bridge Between Image Distributions. In NeurIPS, 2024

    2024

  30. [30]

    Compositional visual generation with energy based models

    Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. InNeurIPS, 2020

    2020

  31. [31]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InICML, 2024

    2024

  32. [32]

    Faster video diffusion with trainable sparse attention

    Peiyuan Zhang, Yongqi Chen, Haofeng Huang, Will Lin, Zhengzhong Liu, Ion Stoica, Eric P Xing, and Hao Zhang. Faster video diffusion with trainable sparse attention. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  33. [33]

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. InNeurIPS, 2024

    2024

  34. [34]

    Nvidia fastgen: Fast generation from diffusion models, 2026

    Weili Nie, Julius Berner, Chao Liu, and Arash Vahdat. Nvidia fastgen: Fast generation from diffusion models, 2026

    2026

  35. [35]

    Lightx2v: Light video generation inference framework.https://github

    LightX2V Contributors. Lightx2v: Light video generation inference framework.https://github. com/ModelTC/lightx2v, 2025

    2025

  36. [36]

    https://github.com/madebyollin/ taehv, 2025

    OllinBoerBohan.Taehv: Tinyautoencoderforhunyuanvideo. https://github.com/madebyollin/ taehv, 2025

    2025

  37. [37]

    Abenchmark dataset and evaluation methodology for video object segmentation

    F.Perazzi,J.Pont-Tuset,B.McWilliams,L.VanGool,M.Gross,andA.Sorkine-Hornung. Abenchmark dataset and evaluation methodology for video object segmentation. InCVPR, 2016

    2016

  38. [38]

    The 2018 davis challenge on video object segmentation.arXiv:1803.00557, 2018

    Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation.arXiv:1803.00557, 2018

    Pith/arXiv arXiv 2018

  39. [39]

    VBench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu,QingyangJin,NattapolChanpaisit,YaohuiWang,XinyuanChen,LiminWang,DahuaLin,YuQiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. InCVPR, 2024

    2024

  40. [40]

    Fastforward: Neural network quantization for research and prototyping

    Qualcomm AI Research. Fastforward: Neural network quantization for research and prototyping. https://github.com/Qualcomm-AI-research/fastforward, 202X

  41. [41]

    Open-sora plan: Open-source large video generation model

    Bin Lin, Yunyang Ge, Xinhua Cheng, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    Pith/arXiv arXiv 2024

  43. [43]

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

    Pith/arXiv arXiv 2024

  44. [44]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong MU, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  45. [45]

    M4v: Multi-modal mamba for text-to-video generation.arXiv preprint arXiv:2506.10915, 2025

    Jiancheng Huang, Gengwei Zhang, Zequn Jie, Siyu Jiao, Yinlong Qian, Ling Chen, Yunchao Wei, and Lin Ma. M4v: Multi-modal mamba for text-to-video generation.arXiv preprint arXiv:2506.10915, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Fast video generation with sliding tile attention

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

    2025

  47. [47]

    Tinyfusion: Diffusion transformers learned shallow

    Gongfan Fang, Kunjun Li, Xinyin Ma, and Xinchao Wang. Tinyfusion: Diffusion transformers learned shallow. InCVPR, 2025

    2025

  48. [48]

    Learnable sparsity for vision generative models

    Yang Zhang, Er Jin, Wenzhong Liang, Yanfei Dong, Ashkan Khakzar, Philip Torr, Johannes Stegmaier, and Kenji Kawaguchi. Learnable sparsity for vision generative models. InICLR, 2026

    2026

  49. [49]

    OBS-diff: Accurate pruning for diffusion models in one-shot

    Junhan Zhu, Hesong Wang, Mingluo Su, Zefang Wang, and Huan Wang. OBS-diff: Accurate pruning for diffusion models in one-shot. InICLR, 2026

    2026

  50. [50]

    Generating long sequences with sparse transformers, 2019

    Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers, 2019

    2019

  51. [51]

    Peters, and Arman Cohan

    Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer, 2020

    2020

  52. [52]

    Bigbird: Transformers for longer sequences

    Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon,PhilipPham,AnirudhRavula,QifanWang,LiYang,andAmrAhmed. Bigbird: Transformers for longer sequences. InNeurIPS, 2020

    2020

  53. [53]

    Efficient content-based sparse attention with routing transformers.Transactions of the Association for Computational Linguistics, 2021

    Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers.Transactions of the Association for Computational Linguistics, 2021

    2021

  54. [54]

    Object-centric diffusion for efficient video editing

    Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. InECCV, 2024

    2024

  55. [55]

    Li, Madian Khabsa, Han Fang, and Hao Ma

    Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020

    2020

  56. [56]

    Reformer: The efficient transformer

    Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. InICLR, 2020

    2020

  57. [57]

    Transformers are rnns: fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: fast autoregressive transformers with linear attention. InICML, 2020

    2020

  58. [58]

    Rethinking attention with performers

    Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. InICLR, 2021

    2021

  59. [59]

    Nyströmformer: A nyström-based algorithm for approximating self-attention

    Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nyströmformer: A nyström-based algorithm for approximating self-attention. InAAAI, 2021

    2021

  60. [60]

    Flashattention: Fast and memory-efficient exact attention with IO-awareness

    Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Re. Flashattention: Fast and memory-efficient exact attention with IO-awareness. InNeurIPS, 2022

    2022

  61. [61]

    Flashattention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. InICLR, 2024

    2024

  62. [62]

    HLA:HadamardLinear Attention, 2026

    HannoAckermann,HongCai,MohsenGhafoorian,andAmirhosseinHabibian. HLA:HadamardLinear Attention, 2026

    2026

  63. [63]

    Gonzalez, Jianfei Chen, and Jun Zhu

    Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, Kaiwen Zheng, Haocheng Xi, Ziteng Wang, Hongzhou Zhu, Min Zhao, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen, and Jun Zhu. SLA: Beyond sparsity in diffusion transformers via fine-tunable sparse–linear attention. InICLR, 2026

    2026

  64. [64]

    Liteattention: A temporal sparse attention for diffusion transformers, 2025

    Dor Shmilovich, Tony Wu, Aviad Dahan, and Yuval Domb. Liteattention: A temporal sparse attention for diffusion transformers, 2025

    2025

  65. [65]

    Adaptor: Adaptive token reduction for video diffusion transformers

    Elia Peruzzo, Adil Karjauv, Nicu Sebe, Amir Ghodrati, and Amir Habibian. Adaptor: Adaptive token reduction for video diffusion transformers. InCVPR, 2025

    2025

  66. [66]

    Astraea: A token-wise acceleration framework for video diffusion transformers

    Haosong Liu, Yuge Cheng, Wenxuan Miao, Zihan Liu, Aiyue Chen, Jing Lin, Yiwu Yao, Chen Chen, Jingwen Leng, Minyi Guo, and Yu Feng. Astraea: A token-wise acceleration framework for video diffusion transformers. InICLR, 2026

    2026

  67. [67]

    TPDiff: Temporal pyramid video diffusion model

    Lingmin Ran and Mike Zheng Shou. TPDiff: Temporal pyramid video diffusion model. InICLR, 2026

    2026

  68. [68]

    PyramidalWan: On making pretrained video model pyramidal for efficient inference

    Denis Korzhenkov, Adil Karjauv, Animesh Karnewar, Mohsen Ghafoorian, and Amirhossein Habibian. PyramidalWan: On making pretrained video model pyramidal for efficient inference. InCVPR, 2026

    2026

  69. [69]

    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, 2024

    2024

  70. [70]

    Align your flow: Scaling continuous-time flow map distillation

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation. InNeurIPS, 2026

    2026

  71. [71]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time consistency distillation. InICCV, 2025

    2025

  72. [72]

    Twinflow: Realizing one-step generation on large models with self-adversarial flows

    Zhenglin Cheng, Peng Sun, Jianguo Li, and Tao Lin. Twinflow: Realizing one-step generation on large models with self-adversarial flows. InICLR, 2026

    2026

  73. [73]

    Metaxas, Sergey Tulyakov, and Jian Ren

    Zhixing Zhang, Yanyu Li, Yushu Wu, yanwu xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris N. Metaxas, Sergey Tulyakov, and Jian Ren. SF-v: Single forward video generation model. InNeurIPS, 2024

    2024

  74. [74]

    Diffusion adversarial post-training for one-step video generation

    Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. InICML, 2025

    2025

  75. [75]

    Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models

    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. InNeurIPS, 2023

    2023

  76. [76]

    Phaseddmd: Few-stepdistributionmatchingdistillationviascorematching within subintervals, 2026

    Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu, Fanzhou Wang, Zhiqian Lin, Tianxiang Ren, Dahua Lin, RuihaoGong,andLeiYang. Phaseddmd: Few-stepdistributionmatchingdistillationviascorematching within subintervals, 2026

    2026

  77. [77]

    MagicDistillation: Weak-to-strong video distillation for large-scale few-step synthesis, 2025

    Shitong Shao, Hongwei Yi, Hanzhong Guo, Tian Ye, Daquan Zhou, Michael Lingelbach, Zhiqiang Xu, and Zeke Xie. MagicDistillation: Weak-to-strong video distillation for large-scale few-step synthesis, 2025

    2025

  78. [78]

    Transition matching distillation for fast video generation, 2026

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

    2026

  79. [79]

    HiStream: Efficient high-resolution video generation via redundancy-eliminated streaming, 2025

    Haonan Qiu, Shikun Liu, Zijian Zhou, Zhaochong An, Weiming Ren, Zhiheng Liu, Jonas Schult, Sen He, Shoufa Chen, Yuren Cong, Tao Xiang, Ziwei Liu, and Juan-Manuel Perez-Rua. HiStream: Efficient high-resolution video generation via redundancy-eliminated streaming, 2025

    2025

  80. [80]

    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. InICLR, 2026

    2026

Showing first 80 references.