Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Plug-and-Play Context Feature Reuse for Efficient Masked Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ReCAP speeds masked image generation up to 2.4x by reusing cached context-token features in lightweight decoding steps, with minimal FID change.

desk verdict A neat KV-cache reuse idea for bidirectional masked generation, with the MAR speedup partly riding on a separate denoising-step reduction that needs an ablation. read the letter →

arxiv 2505.19089 v1 pith:M3U3OWHT submitted 2025-05-25 cs.CV

Xuejie Liu , Anji Liu , Guy Van den Broeck , Yitao Liang This is my paper
classification cs.CV
keywords maskedgenerativemodelsinferenceaccelerationkey-valuecachereusecontextfeaturestabilityimagegenerationplug-and-playdecodingFIDtrade-offNet256
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that masked generative models waste computation by recomputing attention features for every token at every decoding step, even though only a few tokens change between steps. It proposes ReCAP, a plug-and-play module that caches the key-value features of unchanged context tokens and reuses them in cheap partial evaluations interleaved with occasional full evaluations. Across MaskGIT, MAGE, and MAR, ReCAP achieves up to 2.4x faster inference with FID changes around 0.01, and improves the efficiency-fidelity trade-off at every tested step budget. The core claim is that context-token representations are stable enough under small updates that reusing cached features does not accumulate harmful error.

What carries the argument

The central mechanism is the grouped decoding pipeline that alternates Full-FE and Local-FE steps. In a Full-FE, the model computes attention over the entire sequence and caches the key-value pairs of all tokens outside the target subset; in each Local-FE, it recomputes QKV only for the small target subset, concatenates those with the cached KVs, and runs attention, reducing per-step cost from O($N^{2}$) to O(\hat{n}_t N). The stability of context features, measured as cosine similarity of pre-QKV input embeddings before and after token updates, is the empirical justification for trusting the cached KVs across Local-FE steps.

What would settle it

Compute the cosine similarity between the actual cached key and value vectors (rather than pre-QKV embeddings) across consecutive decoding steps for a small context such as K=2 tokens; if this similarity falls well below the near-1 values reported in Figure 2, then cached-KV reuse should produce noticeably worse FID than full recomputation, and ReCAP's speedups would shrink accordingly.

Watch

Extended reading notes

Core claim

The paper's central claim is that the bidirectional attention recomputation in every masked-decoding step is partly wasted work: when a step unmasks only a small subset of tokens, the feature embeddings of the already-decoded context tokens remain nearly unchanged. ReCAP operationalizes this as grouped decoding: each group begins with a full function evaluation (Full-FE) whose key-value pairs for all non-target tokens are cached, followed by one or more local function evaluations (Local-FE) that recompute QKV only for the newly decoded target subset and concatenate those with the cached KVs. The result is a (T + T')-step generation process with only T full-cost steps. On ImageNet256 class-conditional generation, MAR-H+ReCAP at (96+32) steps reaches FID 1.57 while the original MAR-H at 256 steps has FID 1.56, and per-image latency drops from 2.40s to 1.00s; similar trade-off gains appear for MaskGIT and MAGE.

Load-bearing premise

The trick only works if the internal features of already-decoded tokens barely change when a few new tokens are decoded; the paper's support for this measures the inputs to attention, not the actual key-value vectors that ReCAP reuses.

Editorial extensions

If this is right

  • A masked generative model can execute more total decoding steps than full-cost evaluations, so the step-count-versus-fidelity trade-off decouples from raw compute.
  • ReCAP's speedup grows with the number of Local-FEs and with the size of the decoded context, so large models and long-sequence generation benefit most.
  • The same cached-KV scheme applies to encoder-decoder MGMs by caching both encoder and decoder context features, as demonstrated on MAR and MAGE.
  • No retraining or architectural modification is required, so ReCAP can be layered onto future masked generators as their step counts scale.
  • Replacing a subset of full evaluations with Local-FEs preserves FID closely enough that the base model's step-scaling behavior transfers to the accelerated schedule.

Reading between the lines

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

  • Beyond the paper: if the stability trend in Figure 2 persists at higher resolutions or in non-image domains such as language and protein generation, the grouped Full-FE/Local-FE scheme could become a general acceleration primitive for non-autoregressive sequence models.
  • Beyond the paper: the stability analysis measures pre-QKV input embeddings, so a stronger version of the argument would directly measure drift in the cached key and value vectors; if that drift is larger, reuse schedules may need to be more conservative.
  • Beyond the paper: the insertion of Local-FEs could be made adaptive, driven by measured feature drift or a learned policy, which might close the small FID gaps seen in low-step regimes.
  • Beyond the paper: combining outputs from Full-FE and Local-FE steps, rather than discarding the full evaluation's predictions, is a natural extension that could recover some of the remaining fidelity loss.
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

3 major / 5 minor

Summary. The manuscript proposes ReCAP, a training-free module for accelerating masked generative models (MGMs) by interleaving full forward evaluations with lightweight 'Local-FE' steps that cache and reuse key-value (KV) attention features for unchanged context tokens. The method is applied to three MGMs (MaskGIT, MAGE, and MAR) on ImageNet256, covering discrete and continuous token spaces and both conditional and unconditional generation. The central reported result is that MAR-H+ReCAP with (96+32)×2 decoding steps reaches FID 1.57 in 1.00 s per image, versus FID 1.56 in 2.40 s for the original MAR-H at 256×2 steps, and that ReCAP consistently improves speed–fidelity trade-offs.

Significance. If the effect is correctly attributed, ReCAP would be a valuable, simple, and broadly applicable inference-acceleration technique: it requires no retraining, no architectural changes, and it is demonstrated on diverse MGM designs. The paper also contains useful controlled comparisons (MaskGIT-r), a cost breakdown for MAR, and comparisons against strong diffusion and autoregressive baselines. However, a central attribution problem in the MAR experiments must be resolved before the headline speedup claims can be accepted, and the supporting stability analysis is indirect. The core idea is plausible and the requested ablations are feasible within the scope of the manuscript.

major comments (3)
  1. [§6.2, Appendix E.2 (Table 4)] The headline MAR speedups are confounded with a change in the per-token denoising MLP: Appendix E.2 states that 'our ReCAP implementation reduces this to 50 steps for Local-FE' while the base MAR uses 100 denoising MLP steps. For MAR-H, the denoising-MLP time drops from 1.03 s to 0.46 s, a 0.57 s saving that is part of the 1.40 s total time reduction reported in Table 2; for MAR-L the corresponding denoising saving is 0.47 s to 0.14–0.16 s. Because no experiment holds the number of denoising MLP steps fixed and varies only the Full-FE/Local-FE composition, the claim that cached-KV Local-FEs are responsible for the reported 2.4× speedup is not established by the presented data. Please add an ablation that keeps the original 100 denoising MLP steps (or reports separate attention and denoising timings) while applying ReCAP's attention reuse only.
  2. [§5, Figure 2] The stability evidence for the core assumption measures cosine similarity of average-pooled pre-QKV input embeddings for randomly chosen K context tokens, not the actual cached K and V matrices along a real decoding trajectory. Pooling over tokens and layers can hide per-token drift, and the random-context simulation does not mirror the confidence-based or random unmasking order used during generation. Please report per-token KV cosine similarity on actual ReCAP decoding trajectories, or add an experiment that isolates the error from KV reuse by comparing a ReCAP run with a full recomputation of the same sampled tokens under the same schedule.
  3. [§6.1, Appendix A] The MaskGIT-r baseline is not the published MaskGIT model but a reimplementation with substantially hand-tuned sampling hyperparameters (choice temperature τ2(1)=5.5, token temperature schedule in Eq. (2), and a polynomial unmasking schedule replacing the cosine schedule), and the ReCAP hyperparameters u, T, and T′ are chosen per configuration and per model (for example, u=(T+T′)/2 by default for MAR but u=0 for MAGE). The paper should state how these hyperparameters were selected (validation set, grid search, or author expertise) and include a sensitivity analysis; without this, the 'plug-and-play' claim across models is difficult to evaluate and the possibility of favorable selection cannot be excluded.
minor comments (5)
  1. [General] No code or checkpoints are released, and the method depends on several implementation details (grouping, subset partitioning, denoising-MLP-step changes, and sampling schedules); the experiments would be hard to reproduce from the text alone.
  2. [Table 2 and Table 4] FID values are reported as point estimates without variance or confidence intervals; for differences as small as 0.01 (FID 1.56 vs. 1.57), reporting standard errors over multiple seeds or runs would materially strengthen the claim.
  3. [Table 2] The NFE column is ambiguous for ReCAP configurations: it mixes transformer decoding steps, denoising MLP steps, and the ×2 factor from classifier-free guidance. Please define NFE precisely for both baseline and ReCAP rows.
  4. [Figure 3] The color coding for context tokens, masked tokens, and target tokens is difficult to distinguish in grayscale; please add a clear legend and label the cached KVs explicitly.
  5. [§6.2] The sentence 'achieve up to 2∼2.4× faster inference' has a grammatical error and should be rephrased; the same passage would benefit from stating which specific configuration yields the 2× and 2.4× numbers.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: ReCAP's efficiency gains are measured against external FID and timing benchmarks; the only self-citation is related-work context and is not load-bearing, while the denoising-MLP confound is an attribution issue rather than a circular one.

full rationale

ReCAP is an inference-time engineering method evaluated by wall-clock time and FID against the original MaskGIT, MAGE, and MAR models and against external diffusion/autoregressive baselines. The claimed 2.4x speedup is a measured comparison (e.g., MAR-H+ReCAP at (96+32)x2 achieves FID 1.57 in 1.00s versus MAR-H at 256x2 with FID 1.56 in 2.40s, Table 2), not a quantity defined to equal an input. The motivating stability hypothesis in Figure 2 is an empirical observation about feature drift; even though it measures pre-QKV input embeddings rather than the cached KV pairs, that is an evidence-quality weakness, not a circular construction, because the FID results are external validations. The only self-citation is reference [30] in Section 4, cited as related work for the observation that parallel independent updates require many refinement steps in discrete diffusion models; it is not used to justify ReCAP's design or to establish its measured results. Appendix E.2 discloses that ReCAP's MAR configurations also reduce the per-token denoising MLP from 100 to 50 diffusion steps, so the headline speedup is not purely attributable to KV reuse. This is a legitimate correctness/attribution concern, but it is not circularity: both changes are part of the method as evaluated, and the reported FID and time are measured rather than derived. Likewise, per-model hyperparameter choices (u, lt, schedules) affect generality but do not make the evaluation self-justifying. No equation reduces the claimed result to its own assumptions, and no load-bearing claim rests on the authors' prior work. The score of 2 reflects only the presence of a minor, non-load-bearing self-citation, not any circular derivation.

Assumptions & free parameters 4 free parameters · 2 assumptions · 0 invented entities

The central claim rests on the approximation that context token features are stable enough for KV reuse; the free parameters are the configuration choices (u, T', denoising MLP steps) and the MaskGIT-r sampling schedule, all hand-tuned. No invented entities are introduced.

free parameters (4)
  • u (number of initial steps without Local-FEs) = varies (e.g., 10, 12; for MAR default (T+T')/2)
    Chosen per model/configuration to balance speed and quality; affects reported trade-offs.
  • T' (number of Local-FEs) = varies (e.g., 4, 5, 8, 10, 20, 24, 32)
    Set by hand along with T to define each ReCAP configuration; directly controls speedup and quality.
  • Denoising MLP steps for Local-FE in MAR = 50 (vs 100 in original)
    Reduced to cut diffusion time; this is an extra speedup knob that inflates the headline speedup.
  • MaskGIT-r sampling schedule hyperparameters = tau2(1)=5.5; tau_low in {0.65,0.68,0.72,0.75}; polynomial exponent 2.5
    Hand-tuned to make the base MaskGIT scale better with more steps; affects the baseline and comparison.
assumptions (2)
  • standard math Transformer attention mechanics as defined in Vaswani et al. (2017)
    The KV caching and attention formula rely on this standard definition.
  • domain assumption Masked generative models predict tokens from univariate marginals conditioned on context, per MaskGIT/MAR.
    The paper's motivation for the trade-off and its method assume this decoding paradigm.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Plug-and-Play Context Feature Reuse for Efficient Masked Generation." pith.science (2026). https://pith.science/paper/M3U3OWHT

@misc{pith2026250519089,
  author       = {Pith},
  title        = {Pith review of: Plug-and-Play Context Feature Reuse for Efficient Masked Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3U3OWHT}},
  note         = {Machine review of arXiv:2505.19089}
}
read the original abstract

Masked generative models (MGMs) have emerged as a powerful framework for image synthesis, combining parallel decoding with strong bidirectional context modeling. However, generating high-quality samples typically requires many iterative decoding steps, resulting in high inference costs. A straightforward way to speed up generation is by decoding more tokens in each step, thereby reducing the total number of steps. However, when many tokens are decoded simultaneously, the model can only estimate the univariate marginal distributions independently, failing to capture the dependency among them. As a result, reducing the number of steps significantly compromises generation fidelity. In this work, we introduce ReCAP (Reused Context-Aware Prediction), a plug-and-play module that accelerates inference in MGMs by constructing low-cost steps via reusing feature embeddings from previously decoded context tokens. ReCAP interleaves standard full evaluations with lightweight steps that cache and reuse context features, substantially reducing computation while preserving the benefits of fine-grained, iterative generation. We demonstrate its effectiveness on top of three representative MGMs (MaskGIT, MAGE, and MAR), including both discrete and continuous token spaces and covering diverse architectural designs. In particular, on ImageNet256 class-conditional generation, ReCAP achieves up to 2.4x faster inference than the base model with minimal performance drop, and consistently delivers better efficiency-fidelity trade-offs under various generation settings.

Figures

Figures reproduced from arXiv: 2505.19089 by the authors.

Figure 2
Figure 2. Context feature stability during itera￾tive decoding. We measure similarity between con￾text representations before and after token updates, using a pretrained MaskGIT on 50K ImageNet256 samples. At each decoding stage, we extract the input embeddings to the attention module for the K already-decoded tokens. These are average￾pooled within each layer to obtain an aggregated context vector. Cosine similarity is compu… view at source ↗
Figure 3
Figure 3. Grouped Decoding Pipeline with Cached Attention. Inference is organized into T groups, each performing one Full-FE and several Local-FE steps. In the Full-FE, full attention is computed over the entire sequence, and KVs for the static context tokens ( ) and other masked tokens ( ) are cached. In each Local-FE, only the QKVs of the target tokens ( ) are recomputed ( ), while the cached KVs ( ) are reused to form the … view at source ↗
Figure 4
Figure 4. FID vs. inference time for MaskGIT variants and comparative mod￾els. ∗ : taken from the MaskGIT pa￾per [5]. † : with CFG [19]. U-ViT [2] adopts 7 sampling steps in this figure. To assess the impact of ReCAP, we conduct a controlled experiment by fixing the total number of decoding steps T + T ′ and adjusting the allocation between Full- and Local-FEs for ReCAP via the parameter u. As a base￾line, we replace all Loca… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Speed/Performance trade-off for MAR variants and SoTA baselines. ReCAP consistently improves inference efficiency of MAR-Large and -Huge. VARs [48] are SoTA AR models performing next-scale prediction, ∗ denotes the use of KV caching [47]. REPA [56], a SoTA flow-matchin…
Figure 6
Figure 6. Figure 6: FID vs. inference time for un￾conditional generation on ImageNet256. ReCAP consistently achieves lower in￾ference cost across decoding steps, while matching or improving FID. We further evaluate ReCAP on MAGE [27], a state-of￾the-art MGM for unconditional generation wi…
Figure 7
Figure 7. Figure 7: Selected qualitative examples of class-conditional image generation on ImageNet256 using our MAR-L+ReCAP model with (64+20)×2 NFE configuration (FID 1.80, IS 293.9). E.3 Detailed Results of MAGE and MAGE+ReCAP As presented in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 27 canonical work pages

  1. [1]

    Structured denoising diffusion models in discrete state-spaces

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in neural information processing systems , 34:17981–17993, 2021

    work page 2021

  2. [2]

    All are worth words: a vit backbone for score-based diffusion models

    Fan Bao, Chongxuan Li, Yue Cao, and Jun Zhu. All are worth words: a vit backbone for score-based diffusion models. In NeurIPS 2022 Workshop on Score-Based Methods, 2022

    work page 2022

  3. [3]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In International Conference on Learning Representations (ICLR) , 2022

    work page 2022

  4. [4]

    Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan

    Huiwen Chang, Han Zhang, Jarred Barber, Aaron Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to- image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning (ICML) , pages 4055–4075. PMLR, 2023

    work page 2023

  5. [5]

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11315–11325, 2022

    work page 2022

  6. [6]

    Causal diffusion transformers for generative modeling

    Chaorui Deng, Deyao Zhu, Kunchang Li, Shi Guang, and Haoqi Fan. Causal diffusion transformers for generative modeling. arXiv preprint arXiv:2412.12095, 2024

    arXiv 2024

  7. [7]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009

    2009

  8. [8]

    Bert: Pre-training of deep bidi- rectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidi- rectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 4171–4186, 2019

    2019

Show all 57 references
  1. [9]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021

    2021

  2. [10]

    Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis

    Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis. Advances in neural information processing systems, 34:3518–3532, 2021

    2021

  3. [11]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 12873–12883, 2021

    2021

  4. [12]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens

    Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024

    arXiv

  5. [13]

    Mdtv2: Masked diffusion transformer is a strong image synthesizer

    Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023

    2023 arXiv

  6. [14]

    Fully non-autoregressive neural machine translation: Tricks of the trade

    Jiatao Gu and Xiang Kong. Fully non-autoregressive neural machine translation: Tricks of the trade. arXiv preprint arXiv:2012.15833, 2020

    2012 arXiv

  7. [15]

    Vector quantized diffusion model for text-to-image synthesis

    Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10696–10706, 2022

    2022

  8. [16]

    Diffit: Diffusion vision transformers for image generation

    Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. In European Conference on Computer Vision, pages 37–55. Springer, 2024

    2024

  9. [17]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988, 2022

    2022

  10. [18]

    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 NIP, 2017

    2017

  11. [19]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 10

    2022 arXiv

  12. [20]

    Understanding diffusion objectives as the elbo with simple data augmentation

    Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems , 36:65484–65516, 2023

    2023

  13. [21]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models.arXiv preprint arXiv:2404.07724, 2024

    Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models.arXiv preprint arXiv:2404.07724, 2024

    2024 arXiv

  14. [22]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523–11532, 2022

    2022

  15. [23]

    Draft-and-revise: Effective image generation with contextual rq-transformer

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and WOOK SHIN HAN. Draft-and-revise: Effective image generation with contextual rq-transformer. Advances in Neural Information Processing Systems , 35:30127–30138, 2022

    2022

  16. [24]

    Fast inference from transformers via speculative decoding

    Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning , pages 19274–19286. PMLR, 2023

    2023

  17. [25]

    Improved masked image generation with token- critic

    José Lezama, Huiwen Chang, Lu Jiang, and Irfan Essa. Improved masked image generation with token- critic. In European Conference on Computer Vision, pages 70–86. Springer, 2022

    2022

  18. [26]

    Discrete predictor- corrector diffusion models for image synthesis

    Jose Lezama, Tim Salimans, Lu Jiang, Huiwen Chang, Jonathan Ho, and Irfan Essa. Discrete predictor- corrector diffusion models for image synthesis. In The Eleventh International Conference on Learning Representations, 2022

    2022

  19. [27]

    Mage: Masked generative encoder to unify representation learning and image synthesis

    Tianhong Li, Huiwen Chang, Shlok Kumar Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages ...

    2023

  20. [28]

    Return of unconditional generation: A self-supervised representation generation method

    Tianhong Li, Dina Katabi, and Kaiming He. Return of unconditional generation: A self-supervised representation generation method. Advances in Neural Information Processing Systems , 37:125441– 125468, 2024

    2024

  21. [29]

    Autoregressive image generation without vector quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems , 37:56424–56445, 2024

    2024

  22. [30]

    Discrete copula diffusion

    Anji Liu, Oliver Broadrick, Mathias Niepert, and Guy Van den Broeck. Discrete copula diffusion. arXiv preprint arXiv:2410.01949, 2024

    2024 arXiv

  23. [31]

    Maskpredict: Parallel decoding of conditional masked language models

    Marjan Ghazvininejad Omer Levy Yinhan Liu and Luke Zettlemoyer. Maskpredict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing.—-2019, 2019

    2019

  24. [32]

    Discrete diffusion modeling by estimating the ratios of the data distribution

    Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834, 2023

    2023 arXiv

  25. [33]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    2022

  26. [34]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models

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

    2022 arXiv

  27. [35]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, pages 23–40. Springer, 2024

    2024

  28. [36]

    Revisiting non-autoregressive transformers for efficient image synthesis

    Zanlin Ni, Yulin Wang, Renping Zhou, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Shiji Song, Yuan Yao, and Gao Huang. Revisiting non-autoregressive transformers for efficient image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 70...

    2024

  29. [37]

    Adanat: Exploring adaptive policy for token-based image generation

    Zanlin Ni, Yulin Wang, Renping Zhou, Rui Lu, Jiayi Guo, Jinyi Hu, Zhiyuan Liu, Yuan Yao, and Gao Huang. Adanat: Exploring adaptive policy for token-based image generation. In European Conference on Computer Vision, pages 302–319. Springer, 2024

    2024

  30. [38]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  31. [39]

    Scalable diffusion models with Transformers

    William Peebles and Saining Xie. Scalable diffusion models with Transformers. In ICCV, 2023

    2023

  32. [40]

    Strait: Non- autoregressive generation with stratified image transformer

    Shengju Qian, Huiwen Chang, Yuanzhen Li, Zizhao Zhang, Jiaya Jia, and Han Zhang. Strait: Non- autoregressive generation with stratified image transformer. arXiv preprint arXiv:2303.00750, 2023

    2023 arXiv

  33. [41]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. OpenAI, 2018

    2018

  34. [42]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 11

    2019

  35. [43]

    Generating diverse high-fidelity images with vq-vae-2

    Ali Razavi, Aäron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems (NeurIPS) , volume 32, pages 14837–14847, 2019

    2019

  36. [44]

    Beyond next-token: Next-x prediction for autoregressive visual generation

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Beyond next-token: Next-x prediction for autoregressive visual generation. arXiv preprint arXiv:2502.20388, 2025

    2025 arXiv

  37. [45]

    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. In CVPR, 2022

    2022

  38. [46]

    Improved techniques for training GANs

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In NeurIPS, 2016

    2016

  39. [47]

    Fast transformer decoding: One write-head is all you need

    Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019

    1911 arXiv

  40. [48]

    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. Advances in neural information processing systems , 37:84839–84865, 2024

    2024

  41. [49]

    Givt: Generative infinite-vocabulary transformers

    Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292–309. Springer, 2024

    2024

  42. [50]

    Neural discrete representation learning

    Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS) , volume 30, pages 6306–6315, 2017

    2017

  43. [51]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008, 2017

    2017

  44. [52]

    Maskbit: Embedding-free image generation via bit tokens

    Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024

    2024 arXiv

  45. [53]

    Effective and efficient masked image generation models

    Zebin You, Jingyang Ou, Xiaolu Zhang, Jun Hu, Jun Zhou, and Chongxuan Li. Effective and efficient masked image generation models. arXiv preprint arXiv:2503.07197, 2025

    2025

  46. [54]

    Vector-quantized image modeling with improved vqgan.arXiv preprint arXiv:2110.04627, 2021

    Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan.arXiv preprint arXiv:2110.04627, 2021

    2021 arXiv

  47. [55]

    Language model beats diffusion–tokenizer is key to visual generation

    Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023

    arXiv

  48. [56]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024

    arXiv

  49. [57]

    sweet spot

    Yixiu Zhao, Jiaxin Shi, Feng Chen, Shaul Druckmann, Lester Mackey, and Scott Linderman. Informed correctors for discrete diffusion models. arXiv preprint arXiv:2407.21243, 2024. 12 Supplementary Material A Implementation Details of MaskGIT-r First, we adopt the pretrained Mask...

    2024

Pith tools

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