Pith. sign in

REVIEW 6 major objections 4 minor 2 cited by

SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping

T0 review · 6 major / 4 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 The paper claims that two handcrafted frequency features, computed at the 9th autoregressive step, can pick a per-image acceleration strategy that keeps SSIM above a preset threshold while cutting VAR inference time by up to 2.62x on…

desk verdict Useful acceleration idea with a missing ablation that leaves its core adaptive claim unproven. read the letter →

arxiv 2506.08908 v3 pith:PEUKDK5B submitted 2025-06-10 cs.CV

Jiajun Li , Yue Ma , Xinyu Zhang , Qingyan Wei , Songhua Liu , Linfeng Zhang This is my paper
classification cs.CV
keywords visualautoregressivemodelsinferenceaccelerationstepskippingunconditionalbranchreplacementfrequencyanalysissample-adaptivepolicySSIMtext-to-imagegeneration
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

The paper argues that much of the late-stage computation in visual autoregressive (VAR) image generation is redundant, and that the redundant part can be identified per image from two cheap frequency measurements taken early in generation. It proposes SkipVAR, a sample-adaptive accelerator that, at step 9 of a VAR model, computes a Sobel-based high-frequency difference and an FFT-based high-frequency ratio from a downsampled decoded image, then uses a trained logistic-regression decision model to choose one of three policies for the remaining steps: skip them, replace the unconditional branch with the conditional branch, or generate normally. The claim is that this per-image choice preserves fidelity while cutting inference time substantially: over 0.88 average SSIM with up to 1.81x overall acceleration on DrawBench, and up to 2.62x speedup on GenEval with no drop in aggregate score. If right, it means a lightweight, handcrafted-feature policy can beat fixed-ratio token-pruning accelerators without retraining the generative model.

What carries the argument

The load-bearing object is the two-dimensional feature vector (HF_Diff, HF_Ratio) evaluated at the decision step N=9 on a downsampled decoded image, together with the logistic-regression decision model trained to map it to a strategy. HF_Diff measures the inter-step change in Sobel edge energy, capturing whether high-frequency structure has stabilized; HF_Ratio measures the global share of high-frequency Fourier energy beyond a radius, capturing how much fine detail the image inherently contains. The decision model's thresholds convert those two numbers into one of three policy choices for steps 10 through 12, which together account for about 69% of inference time.

What would settle it

Take a content distribution with very different frequency statistics from People-class photos (e.g., dense text pages, low-light photography, or fine line art), apply the step-9 feature extraction and the trained decision model unchanged, and measure SSIM against the unaccelerated model over thousands of samples; if the selected strategies keep SSIM above the chosen threshold no more often than random assignment does, the central claim fails. A second, more local check is the scatter claimed by Figure 7a: compute HF_Ratio at step 9 and HF_Ratio on the final output, and see whether the points still lie on the y=x line across these new distributions.

Watch

Extended reading notes

Core claim

SkipVAR's central claim is that the computational redundancy in VAR inference is well predicted by two measurements of high-frequency content: HF_Diff, the ℓ1 distance between Sobel-filtered decoded images at steps 9 and 8, and HF_Ratio, the fraction of Fourier magnitude beyond a radius in the step-9 decoded image. A logistic-regression model trained on the People class of MJHQ30K maps these two features to the most aggressive acceleration strategy that keeps SSIM against the full model above a preset threshold (0.88, 0.86, or 0.84). Frequency-robust samples are handled by skipping, which stops decoding at an earlier scale; frequency-sensitive samples are handled by unconditional branch replacement, which reuses the conditional output in place of the unconditional branch; both leave the model weights unchanged. The paper reports that this decision rule generalizes across datasets, transfers from the 2B to the 8B Infinity model, and outperforms fixed-ratio token-level methods such as ToMe, SiTo, and FastVAR on the same SSIM/LPIPS comparisons.

Load-bearing premise

The whole method rests on assuming that the Sobel and FFT measurements taken at step 9, on a downsampled decoded image, reliably predict which acceleration strategy will keep a new image close to the full model's output, even for content unlike the People-class training images.

Editorial extensions

If this is right

  • On DrawBench, SkipVAR at the 0.88, 0.86, and 0.84 SSIM thresholds yields 1.58x, 1.70x, and 1.81x speedups respectively while holding SSIM above the chosen threshold and leaving ImageReward and CLIP scores essentially unchanged.
  • On HPSv2, the adaptive policy separates samples: frequency-sensitive images receive the slower unconditional-branch replacement (1.28x, SSIM 0.849), while frequency-robust images receive skipping (1.99x, SSIM 0.905), so the aggregate quality-speed trade-off improves over any single uniform policy.
  • On GenEval, the hand-tuned hybrid schedule (skip steps 11-12, replace unconditional branch at steps 9-10) reaches 2.62x speedup with an overall score of 0.72, slightly above the 0.71 of the unaccelerated Infinity 2B model.
  • Decision models trained on the 2B model transfer to the Infinity-8B model without retraining, giving 1.81x speedup at SSIM 0.901 and 2.06x at SSIM 0.859 on DrawBench.
  • The paper's Appendix D indicates that replacing skipped steps with a super-resolution model (FreqFormer) performs worse than direct skipping, so the gains come from stopping refinement rather than from post-hoc upsampling.

Reading between the lines

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

  • An extension the paper leaves implicit is that the same two features could be computed on latent or intermediate activations instead of decoded, downsampled images, which would move the decision point earlier and reduce the overhead of the decision step itself.
  • Because the policy is trained on SSIM thresholds, it is conservative for images where SSIM falls despite perceptually acceptable output, such as face close-ups; the paper's own Appendix B notes this, and switching the label function to a perceptual metric could unlock more aggressive skipping on exactly the images currently treated as sensitive.
  • The framework treats acceleration as a per-image classification among three discrete policies; a natural extension is to let the decision model output continuous skip/replace budgets, or a ranking over more strategies, rather than one of three options.
  • A ceiling test against an oracle that knows the true SSIM of each strategy would quantify how much of the quality-speed trade-off is lost to the handcrafted features versus to the classifier.
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

6 major / 4 minor

Summary. The paper proposes SkipVAR, a post-training acceleration framework for visual autoregressive (VAR) models that targets two forms of redundancy: redundant later 'high-frequency' refinement steps and redundant unconditional decoding branches. The method introduces two strategies—step skipping and unconditional branch replacement—and uses a lightweight, per-sample decision model based on two handcrafted features (HF_Diff and HF_Ratio) to choose between them at a fixed decision step. Experiments on Infinity-2B and Infinity-8B report SSIM, LPIPS, FID, ImageReward, CLIP, HPSv2, and GenEval results, with headline claims of over 0.88 average SSIM, up to 1.81x overall acceleration, and 2.62x speedup on GenEval.

Significance. If the claims hold, the work would demonstrate that a lightweight per-sample policy based on two handcrafted frequency features can accelerate VAR inference by roughly 1.5 to 2x while keeping outputs close to the full model, and that the policy transfers across model sizes and content categories. The paper is useful to the VAR-acceleration community because it identifies two concrete and well-motivated sources of redundancy, provides multiple benchmark comparisons, and releases code. However, the central adaptive-selection claim is currently supported only indirectly, and the headline numbers are not tied to a single configuration; the significance is therefore conditional on the missing ablations and clarifications.

major comments (6)
  1. [§3.3, Table 1] The sample-adaptive claim is not supported by the current experiments. Section 3.3 describes two acceleration strategies and a decision model that chooses between them, but Table 1 compares only against token-level baselines and a single-feature variant; it never compares SkipVAR with the fixed policies 'always skip', 'always uncond-replace', or a policy using shuffled/random labels at matched speed. Such an ablation is the minimal test of whether the frequency features add predictive value beyond the fixed strategies, and without it the measured speedups could come entirely from the two fixed acceleration recipes. Please add these comparisons, with matched compute budgets and error bars.
  2. [§3.3, Eq. (2), §4.1] The definition of HF_Ratio in Eq. (2) says the high-frequency set H is 'beyond a radius rho (typically rho=0.25)' from the spectrum center, but Section 4.1 states that HF_Ratio is 'thresholded with rho=0.4' for the reported experiments. This inconsistency matters because the decision boundary and downstream conclusions depend on the radius; please state the exact value used for each table and report sensitivity to rho.
  3. [Abstract, Table 1] The headline 'over 0.88 average SSIM with up to 1.81x overall acceleration' is not realized by one operating point in Table 1: +SkipVAR@0.84 achieves 1.81x with SSIM 0.8793, while +SkipVAR@0.88 achieves SSIM 0.9092 at 1.58x. The paper should either present a single configuration achieving both numbers or explicitly rephrase the abstract so the two figures are separate operating points. Similarly, the 2.62x speedup in the abstract and Section 4.2 refers to the 'hybrid (w/o DM)' configuration in Table 4, which is not the adaptive decision-model method and which has SSIM 0.7915 on DrawBench; this needs to be stated clearly.
  4. [§3.3, §4.1] The skip strategy details are incomplete. Section 3.3 says the model 'directly decodes the output at that step without proceeding to further resolution refinements,' but the final output is then at a lower spatial resolution; the paper does not state how this low-resolution image is upsampled to the 1024x1024 reference used for SSIM/LPIPS, nor whether the upsampling time is included in the reported speedups. Appendix D compares super-resolution against direct skipping in an auxiliary experiment, but the main pipeline needs a complete description of the reconstruction and timing protocol.
  5. [Abstract, §3.3, Conclusion] The abstract and conclusion repeatedly describe the method as 'training-free', but Section 3.3 trains logistic regression, decision tree, and random forest classifiers on a 3K-image People dataset. If 'training-free' is meant to mean 'no retraining of the generative model', that should be stated explicitly; as written, the terminology contradicts the method and is likely to mislead readers.
  6. [§3.3, Appendix B] The decision model is labeled by the most aggressive strategy that keeps SSIM above a threshold, and the headline quality metric in Tables 1 and 5 is SSIM to the full model; the authors acknowledge in Appendix B that SSIM may be overly conservative. Because the label source and the main evaluation metric coincide, the main-table numbers are partly self-referential. The held-out Animals accuracy and Infinity-8B transfer mitigate this, but the paper should add at least one evaluation that directly reports the adaptive-vs-fixed comparison using a metric not used in labeling, such as human preference or FID on an independent set.
minor comments (4)
  1. [Section 1] The contribution bullet promises 'extensive experiments and user studies', but no user study is reported anywhere in the manuscript; either add the user-study material or remove the claim.
  2. [Figure 7b] The legend uses abbreviations 'skip_3' and 'uncond_2' that are not defined in the main text; please define them in the caption or in the text around Figure 7.
  3. [Section 4.3] The text reports 'sub-0.03 ms inference latency' for the decision models without describing the hardware or measurement methodology; a brief statement of the timing setup would improve reproducibility.
  4. [Appendix D] Reference [38] (FreqFormer) appears only in Appendix D; please cite it at the first relevant mention in the main text or move the appendix discussion accordingly.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential SSIM labeling: the quality target used to build the decision-model labels is also the headline quality metric, but held-out and external benchmarks keep the central claim largely independent.

  1. fitted input called prediction [Section 3.3 (Labeling and Classifier Training) and Section 4.2 (Comparison on objective metrics), Table 1]
    "Labels are assigned based on the most aggressive acceleration strategy that maintains SSIM above a target threshold, either by skipping steps or replacing unconditional branch with conditional copying. ... On DrawBench (Table 1), SkipVAR maintains SSIM above each model's training threshold and outperforms token-based approaches such as ToMe, SiTo, and FastVAR."

    The decision model is trained on labels that are defined as the most aggressive strategy keeping SSIM above a chosen threshold (0.88, 0.86, or 0.84), and the paper's primary fidelity evidence is then the same SSIM measured against the full model. Reporting that SkipVAR 'maintains SSIM above the training threshold' is therefore partly a restatement of the label-generation rule rather than an independent prediction: on the training distribution the outcome is guaranteed up to classifier error. This is genuine but partial circularity, because the Animals-class evaluation and the Infinity-8B transfer are held out from label construction, and external metrics such as GenEval, HPSv2, ImageReward, CLIP, and FID do not share the SSIM-labeling objective.

full rationale

The paper's derivation chain is mostly self-contained. The two acceleration strategies are defined directly from observed redundancy in VAR inference, and the sample-adaptive selection is a trained classifier on handcrafted Sobel/FFT features; no equation reduces a predicted quantity to a fitted parameter. The one notable self-reference is the SSIM-based labeling loop: labels are chosen as the strategy that keeps SSIM above a threshold, and SSIM against the full model is also the headline quality metric, so the 'quality maintained' claim is partly built into the training objective. However, this is transparent, held-out categories and 8B transfer provide real generalization evidence, and the independent GenEval, HPSv2, ImageReward, and FID results do not depend on the SSIM-labeling procedure. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation. One reporting issue is not circularity: the abstract's 2.62x GenEval speedup comes from '+SkipVAR-hybrid (w/o DM)', a fixed schedule without the adaptive decision model, so it does not by itself validate the sample-adaptive component; that is an attribution concern rather than a circular-derivation concern. Overall the circularity burden is low, so the score is 2.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced. HF_Diff and HF_Ratio are standard image-processing features; the decision model is a trained classifier, not an invented entity. The main burdens are the hand-set decision step, the frequency mask radius, the SSIM thresholds, and the fitted logistic regression weights.

free parameters (4)
  • Decision step N = 9
    Chosen by profiling; acceleration is applied only to steps 10-12. This hand-set design choice determines which features are available and the maximum skip depth.
  • HF_Ratio radius rho = 0.25 in Sec. 3.3, 0.4 in Sec. 4.1
    Radius of the circular mask separating high- and low-frequency components in the FFT domain. The two sections give different values, and no sensitivity analysis is provided.
  • SSIM thresholds = {0.88, 0.86, 0.84}
    User-specified operating points that define strategy labels during training and the reported quality-speed trade-off.
  • Logistic regression weights = Not reported
    Trained on 3,000 MJHQ30K People images with labels derived from SSIM thresholds. These fitted weights are the core of the per-sample decision policy.
assumptions (3)
  • domain assumption The information gap between conditional and unconditional branches is small in later VAR steps, so replacing the unconditional branch with the conditional branch does not change the output much.
    Stated in Sec. 3.2 and Figure 3. This is an empirical observation about Infinity's classifier-free guidance branches, not a derived result, and it underlies the Uncond-Branch-Replace strategy.
  • domain assumption HF_Diff and HF_Ratio computed at step 9 on downsampled decoded images are sufficient predictors of final-image sensitivity to skipped high-frequency steps.
    Introduced in Sec. 3.3. Supported only by Figure 7a and the generalization experiments. If this fails across content distributions, the adaptive policy fails.
  • domain assumption Skipping late steps removes high-frequency refinement only, with no effect on global layout or semantics.
    Used implicitly in Sec. 3.2 to justify evaluating skipped outputs with similarity metrics against the full baseline. The paper's Appendix D compares against super-resolution rather than proving this equivalence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping." pith.science (2026). https://pith.science/paper/PEUKDK5B

@misc{pith2026250608908,
  author       = {Pith},
  title        = {Pith review of: SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEUKDK5B}},
  note         = {Machine review of arXiv:2506.08908}
}
read the original abstract

Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. However, the underlying computational redundancy involved in these steps has yet to be thoroughly investigated. In this paper, we conduct an in-depth analysis of the VAR inference process and identify two primary sources of inefficiency: step redundancy and unconditional branch redundancy. To address step redundancy, we propose an automatic step-skipping strategy that selectively omits unnecessary generation steps to improve efficiency. For unconditional branch redundancy, we observe that the information gap between the conditional and unconditional branches is minimal. Leveraging this insight, we introduce unconditional branch replacement, a technique that bypasses the unconditional branch to reduce computational cost. Notably, we observe that the effectiveness of acceleration strategies varies significantly across different samples. Motivated by this, we propose SkipVAR, a sample-adaptive framework that leverages frequency information to dynamically select the most suitable acceleration strategy for each instance. To evaluate the role of high-frequency information, we introduce high-variation benchmark datasets that test model sensitivity to fine details. Extensive experiments show SkipVAR achieves over 0.88 average SSIM with up to 1.81x overall acceleration and 2.62x speedup on the GenEval benchmark, maintaining model quality. These results confirm the effectiveness of frequency-aware, training-free adaptive acceleration for scalable autoregressive image generation. Our code is available at https://github.com/fakerone-li/SkipVAR and has been publicly released.

Figures

Figures reproduced from arXiv: 2506.08908 by the authors.

Figure 1
Figure 1. Qualitative results of SkipVAR. We show our SkipVAR based-on Infinity-8B [11], demonstrating optimal acceleration (average 1.57× speedup) on samples rich in high-frequency content while preserving visual quality. Abstract Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. H… view at source ↗
Figure 2
Figure 2. SSIM v.s. runtime for different accelera￾tion methods. SkipVAR achieves the best trade-off, maintaining high consistency with faster inference. Based on the VAR framework, Infinity [11] ex￾tends VAR by introducing an infinite-vocabulary tokenizer and bitwise self-correction, enabling high-resolution, photorealistic image generation from text prompts. Recent advances in accelerating visual autoregres￾sive (VAR) model… view at source ↗
Figure 3
Figure 3. Top: L1 loss between conditional and unconditional branches drops over inference steps. Bottom: LPIPS to the final image also drops, with frequency-robust samples consistently lower than frequency-sensitive ones. Motivated by these observations, we propose Skip￾VAR, a sample-adaptive, decision-driven accelera￾tion framework that dynamically selects the most suitable strategy for each instance based on its fre￾quency… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualization of Post-Scaling Steps for Samples with Varying Sensitivity. 2 Related Work Visual Autoregressive Models. Visual Autoregressive (VAR) models [33] depart from traditional token-by-token prediction by adopting a coarse-to-fine, multi-scale next-patch strateg…
Figure 5
Figure 5. Figure 5: Top Left: Overall framework. At the Decision Step, an acceleration strategy is selected and applied to subsequent steps. For clarity in showcasing the acceleration strategies, we omit the propagation of other methods during the acceleration steps. Top Right: Visualizat…
Figure 6
Figure 6. Figure 6: Image generations under different acceleration strategies. Strategies are selected by our decision model. Simpler images use faster strategies with less high-frequency detail; complex ones use slower strategies to preserve more detail. Images are grouped and sorted by …
Figure 7
Figure 7. Figure 7: (a) Scatter of HF_Ratio on the 9th-step downsampled image versus the final high-res output—points lie along y=x, showing reliable low-res Fourier estimates. (b) Logistic regression boundary in HF_Diff–HF_Ratio space for the Animals class. Legend: skip_3 = skip final 3 …
Figure 8
Figure 8. Figure 8: Misjudgments by ImageReward and CLIP Score on HF-Robust (top) and HF-Sensitive [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Token Radius Attention for Efficient Video Generation

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Video diffusion transformers can run ~1.5-2x faster with competitive quality by converting each query's attention entropy into a spatially decayed retention radius instead of dense attention.

  2. FasterVAR: Plug-and-Play Acceleration for Visual Autoregressive Models

    cs.CV 2025-12 conditional novelty 5.0 of 10

    Stage-aware pruning of late generation steps, using random projection and cached-feature restoration, speeds up VAR text-to-image models by up to 3.4x with minimal quality loss.

Reference graph

Works this paper leans on

53 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    GPT-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Token Merging: Your ViT but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token Merging: Your ViT but faster. arXiv preprint arXiv:2210.09461, 2022

    arXiv 2022

  3. [3]

    Token merging for fast stable diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In CVPR, pages 4599–4603, 2023

    work page 2023

  4. [4]

    M-bev: Masked bev perception for robust au- tonomous driving

    Siran Chen, Yue Ma, Yu Qiao, and Yali Wang. M-bev: Masked bev perception for robust au- tonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1183–1191, 2024

    work page 2024

  5. [5]

    Attentive snippet prompting for video retrieval

    Siran Chen, Qinglin Xu, Yue Ma, Yu Qiao, and Yali Wang. Attentive snippet prompting for video retrieval. IEEE Transactions on Multimedia, 26:4348–4359, 2023

    work page 2023

  6. [6]

    Collaborative decoding makes visual auto-regressive modeling efficient

    Zigeng Chen, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Collaborative decoding makes visual auto-regressive modeling efficient. arXiv preprint arXiv:2411.17787, 2024

    arXiv 2024

  7. [7]

    Structural pruning for diffusion models

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In NeurIPS, 2023

    work page 2023

  8. [8]

    Dit4edit: Diffusion transformer for image editing

    Kunyu Feng, Yue Ma, Bingyuan Wang, Chenyang Qi, Haozhe Chen, Qifeng Chen, and Zeyu Wang. Dit4edit: Diffusion transformer for image editing. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 2969–2977, 2025

    work page 2025

Show all 53 references
  1. [9]

    Geneval: An object-focused framework for evaluating text-to-image alignment

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. arXiv preprint arXiv:2310.11513, 2023

    2023 arXiv

  2. [10]

    Fastvar: Linear visual autoregressive modeling via cached token pruning, 2025

    Hang Guo, Yawei Li, Taolin Zhang, Jiangshan Wang, Tao Dai, Shu-Tao Xia, and Luca Benini. Fastvar: Linear visual autoregressive modeling via cached token pruning, 2025. arXiv preprint arXiv:2503.23367v2

    2025 arXiv

  3. [11]

    Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024

    2024 arXiv

  4. [12]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514–7528, 2021

    2021

  5. [13]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), pages 6626–6637, 2017

    2017

  6. [14]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. arXiv preprint arXiv:2402.17245, 2024

    2024 arXiv

  7. [15]

    Faster Diffusion: Rethinking the role of unet encoder in diffusion models

    Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming- Ming Cheng, and Jian Yang. Faster Diffusion: Rethinking the role of unet encoder in diffusion models. NeurIPS, 2023

    2023

  8. [16]

    Autoregressive image generation without vector quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024

    2024 arXiv

  9. [17]

    Q-Diffusion: Quantizing diffusion models

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-Diffusion: Quantizing diffusion models. In ICCV, pages 17535–17545, 2023. 11

    2023

  10. [18]

    Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining

    Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024

    2024 arXiv

  11. [19]

    Faster diffusion via temporal attention decomposition

    Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Faccio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan-Manuel Perez-Rua, and Jürgen Schmidhuber. Faster diffusion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024

    2024 arXiv

  12. [20]

    Avatarartist: Open-domain 4d avatarization

    Hongyu Liu, Xuan Wang, Ziyu Wan, Yue Ma, Jingye Chen, Yanbo Fan, Yujun Shen, Yibing Song, and Qifeng Chen. Avatarartist: Open-domain 4d avatarization. In CVPR, 2025

    2025

  13. [21]

    Learning-to-Cache: Accelerating diffusion transformer via layer caching

    Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-Cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024

    2024 arXiv

  14. [22]

    DeepCache: Accelerating diffusion models for free

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache: Accelerating diffusion models for free. In CVPR, pages 15762–15772, 2024

    2024

  15. [23]

    Follow-your-creation: Empowering 4d creation through video inpainting

    Yue Ma, Kunyu Feng, Xinhua Zhang, Hongyu Liu, David Junhao Zhang, Jinbo Xing, Yinhan Zhang, Ayden Yang, Zeyu Wang, and Qifeng Chen. Follow-your-creation: Empowering 4d creation through video inpainting. arXiv preprint arXiv:2506.04590, 2025

    2025 arXiv

  16. [24]

    Follow your pose: Pose-guided text-to-video generation using pose-free videos, 2024

    Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos, 2024

    2024

  17. [25]

    Follow-your-click: Open-domain regional image animation via short prompts

    Yue Ma, Yingqing He, Hongfa Wang, Andong Wang, Chenyang Qi, Chengfei Cai, Xiu Li, Zhifeng Li, Heung-Yeung Shum, Wei Liu, et al. Follow-your-click: Open-domain regional image animation via short prompts. arXiv preprint arXiv:2403.08268, 2024

    2024 arXiv

  18. [26]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait animation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024

    2024

  19. [27]

    Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning

    Yue Ma, Yulong Liu, Qiyuan Zhu, Ayden Yang, Kunyu Feng, Xinhua Zhang, Zhifeng Li, Sirui Han, Chenyang Qi, and Qifeng Chen. Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207, 2025

    2025

  20. [28]

    Visual knowledge graph for human action reasoning in videos

    Yue Ma, Yali Wang, Yue Wu, Ziyu Lyu, Siran Chen, Xiu Li, and Yu Qiao. Visual knowledge graph for human action reasoning in videos. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4132–4141, 2022

    2022

  21. [29]

    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. In CVPR, pages 14297–14306, 2023

    2023

  22. [30]

    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

    2022 arXiv

  23. [31]

    Post-training quantization on diffusion models

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In CVPR, pages 1972–1981, 2023

    1972

  24. [32]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024

    2024 arXiv

  25. [33]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024

    2024

  26. [34]

    Unipaint: Unified space-time video inpainting via mixture-of-experts

    Zhen Wan, Yue Ma, Chenyang Qi, Zhiheng Liu, and Tao Gui. Unipaint: Unified space-time video inpainting via mixture-of-experts. arXiv preprint arXiv:2412.06340, 2024

    2024 arXiv

  27. [35]

    Attention- driven training-free efficiency enhancement of diffusion models

    Hongjie Wang, Difan Liu, Yan Kang, Yijun Li, Zhe Lin, Niraj K Jha, and Yuchen Liu. Attention- driven training-free efficiency enhancement of diffusion models. In CVPR, pages 16080–16089, 2024. 12

    2024

  28. [36]

    Cove: Un- leashing the diffusion feature correspondence for consistent video editing

    Jiangshan Wang, Yue Ma, Jiayi Guo, Yicheng Xiao, Gao Huang, and Xiu Li. Cove: Un- leashing the diffusion feature correspondence for consistent video editing. arXiv preprint arXiv:2406.08850, 2024

    2024 arXiv

  29. [37]

    Taming rectified flow for inversion and editing

    Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024

    2024 arXiv

  30. [38]

    Freqformer: Frequency-domain transformer for 3-d visualization and quantification of human retinal circula- tion

    Lingyun Wang, Bingjie Wang, Jay Chhablani, Jose Alain Sahel, and Shaohua Pi. Freqformer: Frequency-domain transformer for 3-d visualization and quantification of human retinal circula- tion. arXiv preprint arXiv:2411.11189, 2024

    2024

  31. [39]

    EMU3: Next-token prediction is all you need

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. EMU3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024

    2024 arXiv

  32. [40]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600– 612, 2004

    2004

  33. [41]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023

    2023 arXiv

  34. [42]

    Show-o: One single trans- former to unify multimodal understanding and generation

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single trans- former to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024

    2024 arXiv

  35. [43]

    Litevar: Compressing visual autoregressive modelling with efficient attention and quantization

    Rui Xie, Tianchen Zhao, Zhihang Yuan, Rui Wan, Wenxi Gao, Zhenhua Zhu, Xuefei Ning, and Yu Wang. Litevar: Compressing visual autoregressive modelling with efficient attention and quantization. arXiv preprint arXiv:2411.17178, 2024

    2024 arXiv

  36. [44]

    Enhancing image generation fidelity via progressive prompts

    Zhen Xiong, Yuqi Li, Chuanguang Yang, Tiao Tan, Zhihong Zhu, Siyuan Li, and Yue Ma. Enhancing image generation fidelity via progressive prompts. arXiv preprint arXiv:2501.07070, 2025

    2025 arXiv

  37. [45]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, 2023

    2023

  38. [46]

    Eedit: Rethinking the spatial and temporal redundancy for efficient image editing

    Zexuan Yan, Yue Ma, Chang Zou, Wenteng Chen, Qifeng Chen, and Linfeng Zhang. Eedit: Rethinking the spatial and temporal redundancy for efficient image editing. arXiv preprint arXiv:2503.10270, 2025

    2025

  39. [47]

    Follow- your-multipose: Tuning-free multi-character text-to-video generation via pose guidance

    Beiyuan Zhang, Yue Ma, Chunlei Fu, Xinyang Song, Zhenan Sun, and Ziqiang Li. Follow- your-multipose: Tuning-free multi-character text-to-video generation via pose guidance. arXiv preprint arXiv:2412.16495, 2024

    2024 arXiv

  40. [48]

    Token pruning for caching better: 9 times acceleration on stable diffusion for free

    Evelyn Zhang, Bang Xiao, Jiayi Tang, Qianli Ma, Chang Zou, Xuefei Ning, Xuming Hu, and Linfeng Zhang. Token pruning for caching better: 9 times acceleration on stable diffusion for free. arXiv preprint arXiv:2501.00375, 2024

    2024 arXiv

  41. [49]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

    2018

  42. [50]

    Magiccolor: Multi- instance sketch colorization

    Yinhan Zhang, Yue Ma, Bingyuan Wang, Qifeng Chen, and Zeyu Wang. Magiccolor: Multi- instance sketch colorization. arXiv preprint arXiv:2503.16948, 2025

    2025 arXiv

  43. [51]

    Multibooth: Towards generating all your concepts in an image from text

    Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Xiu Li. Multibooth: Towards generating all your concepts in an image from text. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 10923–10931, 2025. 13

    2025

  44. [52]

    Instantswap: Fast customized concept swapping across sharp shape differences

    Chenyang Zhu, Kai Li, Yue Ma, Longxiang Tang, Chengyu Fang, Chubin Chen, Qifeng Chen, and Xiu Li. Instantswap: Fast customized concept swapping across sharp shape differences. arXiv preprint arXiv:2412.01197, 2024

    2024 arXiv

  45. [53]

    sensitive

    Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffusion transformers with token-wise feature caching. arXiv preprint arXiv:2410.05317, 2024. 14 A Scalability of Acceleration Strategies. Although our current acceleration strategy uses skip or un...

    2024 arXiv

Pith tools

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