Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Bag of Design Choices for Inference of High-Resolution Masked Generative Transformer

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A set of inference-time design choices lets masked generative transformers (MGTs) beat their own vanilla sampling about 70% of the time on human-preference benchmarks, with no retraining, and cuts memory from 11.98 to 4.57 GB via…

desk verdict A genuinely useful bag of inference-time tricks for masked generative transformers, but the headline win rate is measured on the same 150-prompt subset used for tuning, so treat it as promising rather than established. read the letter →

arxiv 2411.10781 v2 pith:OIZT2SCB submitted 2024-11-16 cs.CV cs.LG

Shitong Shao , Zikai Zhou , Tian Ye , Lichen Bai , Zhiqiang Xu , Zeke Xie This is my paper · ORCID
classification cs.CVcs.LG
keywords maskedgenerativetransformerinference-timedesignchoicestext-to-imagegenerationnoisescheduleZ-Samplingquantizationtokenmergingdifferentialsampling
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 asks whether a masked generative Transformer (MGT) - a model that generates images by repeatedly predicting randomly masked visual tokens - can be made better at inference time without retraining. It claims yes: a collection of small changes to the sampling loop beats vanilla MaskGIT and Meissonic sampling, with combined methods winning about 70% of human-preference comparisons on HPS v2 at 1024x1024. It also claims memory can be cut from 11.98 GB to 4.57 GB with negligible quality loss using its quantization scheme. The paper's contribution is a set of tested, composable inference-side techniques plus a comparison showing which diffusion-model tricks transfer to MGT and which need redesign.

What carries the argument

The object is the inference loop of an MGT: start with all tokens masked, at each step predict all masked tokens, replace them, then re-mask the lowest-confidence predictions, and repeat. The paper treats this loop as a set of replaceable design choices, each targeting one stage: the noise schedule controls how many tokens are masked at each step; masked Z-Sampling inserts a forward-then-backtrack pass using low-confidence masking; noise regularization and differential sampling reshape the categorical distribution before token selection; SCQ compresses the Transformer; TomeMGT merges tokens in multi-modal Transformer blocks; and a momentum-based solver mixes tokens from two steps in proportion to the schedule. The identity doing the work is the schedule together with confidence-based masking: everything hangs on the model's self-reported token probabilities.

What would settle it

Take a masked generative Transformer trained with the standard cosine mask schedule, sample the same prompts under cosine and under $1 - t^{0.6}$ with identical seeds and CFG settings, and compare human-preference or FID scores across several unrelated models; the noise-schedule claim is falsified if the gains do not reproduce or reverse sign. Similarly, running SCQ on another MGT and checking whether memory drops while quality holds would test the quantization claim.

Watch

Extended reading notes

Core claim

The central empirical discovery is that the standard MGT inference loop is suboptimal in several independent ways. First, replacing the cosine mask schedule with $1 - t^\rho$ at inference, with $\rho$ near 0.6, improves quality even though the model was trained with cosine. Second, Z-Sampling works only when the backtracking step remasks low-confidence predicted tokens rather than random tokens. Third, adding timestep-dependent Gaussian noise to the Transformer logits (noise regularization) and resampling tokens whose neighboring distributions are too similar (differential sampling) both improve output quality and diversity. Fourth, quantizing weights to 4 bits and only low-amplitude activations to 8 bits (SCQ) preserves quality while cutting memory by more than half. The methods compose: combined they give about a 70% winning rate over vanilla sampling on HPS v2 with Meissonic-1024.

Load-bearing premise

The paper assumes an MGT trained with the cosine schedule still produces trustworthy token probabilities when sampled under the different schedule $1 - t^\rho$, with no retraining or alignment, and that the same holds for the momentum-based solver's ODE-style token mixing.

Editorial extensions

If this is right

  • MGT quality can be improved at inference time with zero retraining by replacing the cosine mask schedule with $1 - t^\rho$ near $\rho = 0.6$.
  • Diffusion-derived sampling tricks transfer to MGT only after redesign: random Z-Sampling hurts, but low-confidence masked Z-Sampling helps.
  • Injecting stochasticity into the logits late in sampling helps MGT, in contrast to diffusion models where it typically hurts.
  • SCQ lowers peak memory from 11.98 GB to 4.57 GB on Meissonic-1024 with minimal metric change, enabling lower-memory deployment.
  • The design choices compose, reaching about 70% win rate over vanilla sampling on HPS v2, and they also transfer to MaskGIT and to the autoregressive LlamaGen.

Reading between the lines

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

  • If the schedule-mismatch result generalizes, MGT training could be made robust to a family of inference schedules, potentially turning the noise schedule into a free per-prompt knob rather than a fixed training choice.
  • The success of late-timestep noise regularization suggests MGT token distributions become overconfident near the end of sampling; an explicit confidence-calibration method might achieve the same gains more directly.
  • SCQ's heuristic of quantizing only low-amplitude activations could be combined with per-layer sensitivity analysis, and its memory savings would be larger on models with more tokens than Meissonic's 1024.
  • Differential sampling's KL-based resampling is architecture-agnostic; a similar redundancy-rejection rule could apply to speculative decoding or multi-token prediction in language models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper studies inference-time design choices for masked generative transformers (MGTs), targeting high-resolution text-to-image models such as Meissonic and MaskGIT. The authors propose and evaluate several training-free modifications: a Karras-style noise schedule 1-t^rho replacing the cosine schedule, masked Z-Sampling (a rescheduling technique that masks low-confidence tokens when backtracking), noise regularization that perturbs logits before softmax with timestep-dependent variance I(t), and differential sampling that resamples tokens whose adjacent-step KL divergence is low. They also investigate efficiency methods: SCQ (W4A8 quantization with secondary calibration of low-magnitude activations), TomeMGT (token merging adapted to MGT), and a momentum-based solver inspired by DPM-Solver. The headline claim is that combining the enhanced-inference design choices achieves winning rates of approximately 70% over vanilla sampling on HPS v2 with Meissonic-1024x1024, and that SCQ reduces total memory from 11.98 GB to 4.57 GB with minimal performance loss. Additional evaluations are reported on GenEval, T2I-Compbench, HPD v2 subsets, ChallengeBench, MaskGIT, and LlamaGen.

Significance. If the empirical claims are robust, the paper would provide a useful, training-free toolbox for improving MGT inference quality and memory efficiency, with potential practical impact for high-resolution text-to-image generation. The paper has several strengths: it reports negative or limited results honestly (deterministic sampling, TomeMGT's marginal gains, momentum solver's failure at large N), evaluates on multiple benchmarks and metrics, includes a newly constructed ChallengeBench, and tests transfer to MaskGIT and LlamaGen. The SCQ memory reduction is concrete and well supported by Table 3. However, the central quantitative claim currently rests on a 150-prompt subset that was also used for hyperparameter selection, and the combined configuration is not evaluated on a held-out set or on the full HPD v2 benchmark; no error bars or multiple-seed analyses are provided. The strength of the contribution therefore depends on the evaluation protocol being strengthened.

major comments (4)
  1. [Section 5, Fig. 10, Table 16] The headline claim of approximately 70% winning rates is computed on the HPD v2 Subset (150 prompts, Appendix 7.1), which is the same set used to select rho, z, I(t), and the inversion CFG scale. Appendix 9.1 explicitly acknowledges 'the potential risk of overfitting introduced by this exhaustive search approach on the hyperparameter rho,' but the full HPD v2 results in Tables 9, 12, and 13 are reported only for individual methods, not for the combined configuration that yields the headline number. The central claim that the design choices can be merged therefore lacks held-out validation. I request a freeze-then-evaluate protocol: either a held-out split of HPD v2, or full-benchmark evaluation of the combined configuration, with winning rates reported per metric and domain.
  2. [Section 3.2, Fig. 6 (Left)] The ablation of the inversion CFG scale shows the best average performance near -1 and 9, yet the authors set the inversion CFG scale to 0 and the standard CFG scale to 9 for all subsequent experiments, citing computational cost. This means the operating point used in the downstream experiments, including the headline result, is not the ablated optimum on the tuning set. The paper should either justify the suboptimal operating point on the held-out evaluation of the combined configuration, or move to the ablated optimum and re-evaluate; the current choice makes it difficult to separate the effect of masked Z-Sampling from the effect of choosing a non-optimal inversion CFG scale.
  3. [Section 3.4, Fig. 9; Section 3.3, Fig. 7] The hyperparameters z=75 for differential sampling and I(t)=|cos(pi t)| for noise regularization are selected on the same 150-prompt HPD v2 Subset used for the headline combined evaluation. Beyond the selection-on-evaluation concern, no interaction analysis is reported: the combination of all three enhanced-inference methods (noise regularization, differential sampling, masked Z-Sampling) is evaluated only at the individually selected operating points. Please report a small interaction study (e.g., z in {50, 75, 100} and I(t) alternatives) for the combined configuration, or at least provide full-benchmark results for the combined configuration so that interactions can be assessed independently of the tuning set.
  4. [Figures 4, 6, 7, 9, 10, and Tables 2, 3, 15, 16] No error bars, confidence intervals, or multiple-run statistics are reported anywhere in the paper. All comparisons are point estimates of metric means or winning rates over 150 prompts (or four 800-prompt domains). A 70% winning rate over 150 paired comparisons carries a wide confidence interval; without a variance estimate, the reported gains over vanilla sampling cannot be distinguished from chance variation under hyperparameter selection. I request bootstrap confidence intervals for the winning rates and, where feasible, multiple sampling seeds.
minor comments (5)
  1. [Section 3.2, Fig. 6 (Left)] The x-axis of Fig. 6 (Left) ranges from 0 to 12, but the text states that the best values are near -1 and 9; please extend the axis to include negative values or clarify how negative inversion CFG scales were evaluated.
  2. [Throughout the main text] The term 'HPD v2' is used ambiguously: Section 3 states that experiments use the HPD v2 Subset (150 prompts, Appendix 7), while Section 5 and Fig. 10 refer to 'HPD v2' without qualification. Please consistently distinguish 'HPD v2 Subset' from the full 'HPD v2' benchmark.
  3. [Section 4.3, Fig. 11] The momentum-based solver is described as an acceleration approach but underperforms vanilla sampling for large N; the discussion should state explicitly that this component is not part of the combined enhanced-inference configuration behind the headline result, so that readers do not infer that all listed methods contribute to the 70% claim.
  4. [Table 3] The SCQ description in Section 4.1 says one-third of activation values are quantized by default, but no ablation is shown for this fraction; please provide a sensitivity analysis or justify the fraction, and report the CPU-offloading overhead (transfer time and host memory) alongside the GPU memory reduction.
  5. [Section 3.3, Eq. (1)] The symbol I(t) is used both for the identity matrix in the Gaussian noise notation and for the timestep-dependent standard deviation function; please use distinct notation (e.g., sigma_t for the standard deviation) to avoid confusion.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline ~70% winning rate is computed on the same 150-prompt HPD v2 Subset used to tune rho, z%, I(t), and inversion CFG scale; the paper itself concedes the overfitting risk for rho.

  1. fitted input called prediction [Sec. 3.1 (Table 2) and Sec. 5 (Fig. 10); Appendix 9.1]
    "All experiments, unless otherwise specified, were conducted with Meissonic-1024×1024 on the HPD v2 Subset (see Appendix 7). ... Given this observation, we sample ρ more densely and uniformly within the interval [0.5, 1] to obtain more precise results ... When ρ = 0.6, 1 − tρ demonstrates more favorable behavior compared to the standard cosine schedule cos( πt 2 ). ... Specifically, when integrated with the other two strategies, masked Z-Sampling achieves winning rates of approximately 70% compared to vanilla sampling on HPD v2. ..."

    The ρ value is obtained by exhaustive search on the HPD v2 Subset, which is the default experiment set, and then Table 2 on that same subset is used as evidence that 1−t^ρ is better than cosine; the 'better schedule' conclusion is the argmax of the search, so the comparison is fitted rather than independently predicted. The same default subset also underlies the ~70% winning-rate headline for the merged configuration, whose other hyperparameters (z%=75, I(t)=|cos(πt)|, inversion CFG scale) were likewise ablated on that subset. The paper's own Appendix 9.1 concedes the overfitting risk for ρ. Independent support exists on other benchmarks, but the headline number is not an out-of-sample test of the design choices.

full rationale

Most of this paper is an empirical bag of design choices rather than a formal derivation chain: masked Z-Sampling, noise regularization, differential sampling, SCQ, TomeMGT, and the momentum-based solver are compared against vanilla sampling on external benchmarks (GenEval, T2I-Compbench, MaskGIT IS/FID), and the SCQ memory claim is a direct torch.cuda.max_memory_reserved measurement. I found no self-definitional equation, no imported uniqueness theorem, and no load-bearing self-citation chain; the only author-overlap citation (Meissonic) is used as a base model, not as evidence for the new methods. The substantive issue is that the headline ~70% winning rate and the schedule/differential-sampling findings come from exhaustive hyperparameter search on the 150-prompt HPD v2 Subset and are then reported on that same default subset, with Appendix 9.1 explicitly acknowledging the overfitting risk for rho. This makes the central quantitative claim partially a fitted result rather than an independent prediction. Because individual methods also receive support on other benchmarks and the paper is otherwise self-contained, the appropriate score is 4 rather than 6 or higher.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the MGT sampling-as-denoising analogy, on several domain-specific heuristics about token confidence and diversity, and on the representativeness of the evaluation subsets. The free parameters are mostly hyperparameters selected on the same benchmark used for the headline results, which is the main source of uncertainty.

free parameters (5)
  • rho in noise schedule 1 - t^rho = 0.6 (optimal on HPD v2 Subset; searched over [0.015625, 64])
    Section 3.1 and Table 2; the authors explicitly acknowledge overfitting risk from exhaustive search in Appendix 9.1.
  • z% in differential sampling = 75 (best on HPD v2 Subset; z in [0, 100])
    Section 3.4 and Figure 9; chosen by ablation on the same evaluation subset used for the headline comparisons.
  • noise schedule I(t) for noise regularization = |cos(pi t)| (selected among tested curves)
    Section 3.3 and Figure 7; selected by comparing winning rates against vanilla sampling on the HPD v2 Subset.
  • inversion CFG scale for masked Z-Sampling = 0 in main experiments; ablations showed local optima near -1 and 9
    Section 3.2 and Figure 6 Left; set to 0 to avoid additional NFEs, but the choice is informed by the same ablation data.
  • quantized activation fraction in SCQ = one-third of activation values
    Section 4.1; described as the default choice, with no principled selection criterion or sensitivity analysis.
assumptions (6)
  • domain assumption Inference-time noise schedule can be changed from the training schedule without invalidating the model's token predictions.
    Section 3.1 explores 1 - t^rho and claims better performance despite training with the cosine schedule.
  • domain assumption Low-confidence predicted tokens are safe to mask during Z-Sampling backtracking because they carry less critical semantic content.
    Section 3.2 and Figure 5; this motivates masked Z-Sampling over random masking.
  • domain assumption KL divergence between adjacent-step output distributions identifies redundant tokens whose resampling improves quality.
    Section 3.4 and Equation 3; the entire differential sampling method rests on this heuristic.
  • domain assumption Adding Gaussian noise to logits before softmax increases diversity without degrading semantic alignment.
    Section 3.3 and Equation 1; noise regularization is justified empirically rather than theoretically.
  • domain assumption DPM-Solver flow-matching equations apply to discrete MGT sampling, and token-proportion merging approximates the arithmetic operations.
    Section 4.3 and Appendix 9.8; the momentum solver transforms ODE arithmetic into discrete token replacement.
  • domain assumption The HPD v2 Subset of 150 randomly selected prompts is representative enough for tuning and comparing inference methods.
    Section 7.1 defines this subset; most headline numbers and hyperparameter choices in the main text rely on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bag of Design Choices for Inference of High-Resolution Masked Generative Transformer." pith.science (2026). https://pith.science/paper/OIZT2SCB

@misc{pith2026241110781,
  author       = {Pith},
  title        = {Pith review of: Bag of Design Choices for Inference of High-Resolution Masked Generative Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIZT2SCB}},
  note         = {Machine review of arXiv:2411.10781}
}
read the original abstract

Text-to-image diffusion models (DMs) develop at an unprecedented pace, supported by thorough theoretical exploration and empirical analysis. Unfortunately, the discrepancy between DMs and autoregressive models (ARMs) complicates the path toward achieving the goal of unified vision and language generation. Recently, the masked generative Transformer (MGT) serves as a promising intermediary between DM and ARM by predicting randomly masked image tokens (i.e., masked image modeling), combining the efficiency of DM with the discrete token nature of ARM. However, we find that the comprehensive analyses regarding the inference for MGT are virtually non-existent, and thus we aim to present positive design choices to fill this gap. We propose and redesign a set of enhanced inference techniques tailored for MGT, providing a detailed analysis of their performance. Additionally, we explore several DM-based approaches aimed at accelerating the sampling process on MGT. Extensive experiments and empirical analyses on the recent SOTA MGT, such as MaskGIT and Meissonic lead to concrete and effective design choices, and these design choices can be merged to achieve further performance gains. For instance, in terms of enhanced inference, we achieve winning rates of approximately 70% compared to vanilla sampling on HPS v2 with Meissonic-1024x1024.

Figures

Figures reproduced from arXiv: 2411.10781 by the authors.

Figure 1
Figure 1. Visualization of our design choices on Meissonic-512 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The complete sampling pipeline of MaskGIT and Meissonic, and how our proposed effective and specific design choices [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of different noise schedules. The black [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Visualization of the performance of different noise [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Left: Ablation study on the inversion CFG scale. Right: Comparison between masked Z-Sampling and vanilla sampling on the HPD v2 dataset. We present the ablation results in [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Trajectory of KL divergence between distributions at [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 7
Figure 7. Figure 7: Ablation study on I(t). Note that 1 2 | cos(πt)|−| cos(πt)| and | cos(πt)| − 1 2 | cos(πt)| represent the scenarios where the function 1 2 | cos(πt)| or | cos(πt)| is used for t ∈ [0, 1 2 ), while | cos(πt)| or 1 2 | cos(πt)| is used for t ∈ [ 1 2 , 1], respectively. I…
Figure 9
Figure 9. Figure 9: Ablation Studies of z% in differential sampling. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]
Figure 10
Figure 10. Figure 10: Comparison of the combination of our methods on HPD v2. The results for Meissonic-512 [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: Ablation Studies of Momentum-based Solver. [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: The visualization of deterministic sampling, where the number following [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 15
Figure 15. Figure 15: Left: The sampling trajectory of noise regular￾ization. Different colored tokens represent different values ∈ {0, · · · , 8191}. Right: The visualization of the entropy as the sampling progresses. Shaded blocks are the standard deviations. Method PickScore (↑) HPS v2 …
Figure 14
Figure 14. Figure 14: Visualization of the performance of different noise [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection

    cs.CV 2024-12 conditional novelty 6.0 of 10

    Z-Sampling alternates high-guidance denoising and low-guidance inversion at each step to improve prompt alignment in pretrained text-to-image diffusion models.

Reference graph

Works this paper leans on

57 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

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

    arXiv

  2. [2]

    Zigzag diffusion sampling: The path to success ls zigzag

    Anonymous. Zigzag diffusion sampling: The path to success ls zigzag. In Submitted to The Thirteenth International Con- ference on Learning Representations, 2024. under review. 2, 4, 5

    work page 2024

  3. [3]

    Meissonic: Revitalizing masked generative trans- formers for efficient high-resolution text-to-image synthesis

    Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing masked generative trans- formers for efficient high-resolution text-to-image synthesis. arXiv preprint arXiv:2410.08261, 2024. 1, 3, 4

    arXiv

  4. [4]

    Token merging for fast sta- ble diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast sta- ble diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599–4603,

  5. [5]

    Maskgit: Masked generative image transformer

    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, pages 11315–11325, 2022. 1, 3, 4, 8

    work page 2022

  6. [6]

    Muse: Text-to-image generation via masked generative transform- ers

    Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Mur- phy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transform- ers. arXiv preprint arXiv:2301.00704, 2023. 1, 3

    arXiv 2023

  7. [7]

    Q-dit: Ac- curate post-training quantization for diffusion transformers

    Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Ac- curate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. 4

    arXiv 2024

  8. [8]

    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. 1

    arXiv

Show all 57 references
  1. [9]

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

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text- to-image alignment. Advances in Neural Information Pro- cessing Systems, 36:52132–52152, 2023. 7, 8

    2023

  2. [10]

    Clipscore: A reference-free evaluation met- ric for image captioning, 2022

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning, 2022. 2

    2022

  3. [11]

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

    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 equi- librium. In Neural Information Processing Systems , Long Beach Convention Center, Long Beach, 2017. NeurIPS. 3, 1

    2017

  4. [12]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In Neural Information Processing Systems Work- shop, Virtual Event, 2021. NeurIPS. 4

    2021

  5. [13]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In Neural Information Processing Systems, pages 6840–6851, Virtual Event, 2020. NeurIPS. 1, 3, 4

    2020

  6. [14]

    T2i-compbench: A comprehensive bench- mark for open-world compositional text-to-image genera- tion

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive bench- mark for open-world compositional text-to-image genera- tion. Advances in Neural Information Processing Systems , 36:78723–78747, 2023. 7, 8

    2023

  7. [15]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceed- ings of the IEEE conference on computer vision ...

    2018

  8. [16]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565–26577, 2022. 2, 4, 5

    2022

  9. [17]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4

    2013 arXiv

  10. [18]

    Black Forest Labs. Flux. https : / / blackforestlabs.ai/, 2024. 1, 7, 5

    2024

  11. [19]

    Laion-aesthetics

    Laion.ai. Laion-aesthetics. https://laion.ai/blog/ laion-aesthetics/, 2022. 3, 1

    2022

  12. [20]

    Minimizing trajectory curvature of ode-based generative models

    Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. arXiv preprint arXiv:2301.12003, 2023. 5

    2023 arXiv

  13. [21]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer, 2014. 3, 1

    2014

  14. [22]

    Alignment of dif- fusion models: Fundamentals, challenges, and future

    Buhua Liu, Shitong Shao, Bao Li, Lichen Bai, Haoyi Xiong, James Kwok, Sumi Helal, and Zeke Xie. Alignment of dif- fusion models: Fundamentals, challenges, and future. arXiv preprint arXiv:2409.07253, 2024. 1

    2024

  15. [23]

    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. 1, 3

    2024 arXiv

  16. [24]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 4, 5

    2022 arXiv

  17. [25]

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

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, and Chongx- uan Li. Dpm-solver++: Fast solver for guided sam- pling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 7

    2022 arXiv

  18. [26]

    Dpm-solver: A fast ode solver for diffusion 9 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 9 probabilistic model sampling in around 10 steps. In Neural Information Processing Systems , New Orleans, LA, USA,

  19. [27]

    On distillation of guided diffusion models

    Chenlin Meng, Ruiqi Gao, Diederik P Kingma, Stefano Er- mon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. arXiv preprint arXiv:2210.03142,

    arXiv

  20. [28]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2

    2023

  21. [29]

    Learning to reason with llms, 2024

    OpenAI. Learning to reason with llms, 2024. 1

    2024

  22. [30]

    Scalable diffusion models with transformers

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

  23. [31]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth Interna- tional Conference on Learning Representations. 1, 5

  24. [32]

    Not all noises are created equally: Diffusion noise selection and optimization

    Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041, 2024. 7

    2024 arXiv

  25. [33]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 1

    2021

  26. [34]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Con- ference on Learning Representations , Virtual Event, 2022. OpenReview.net. 8

    2022

  27. [35]

    Improved tech- niques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved tech- niques for training gans. In Neural Information Process- ing Systems , Centre Convencions Internacional Barcelona, Barcelona SPAIN, 2016. NeurIPS. 3, 2

    2016

  28. [36]

    Neural machine translation of rare words with subword units

    Rico Sennrich. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 1

    2015 arXiv

  29. [37]

    Catch-up distillation: You only need to train once for accelerating sampling

    Shitong Shao, Xu Dai, Shouyi Yin, Lujun Li, Huanran Chen, and Yang Hu. Catch-up distillation: You only need to train once for accelerating sampling. arXiv preprint arXiv:2305.10769, 2023. 1

    2023 arXiv

  30. [38]

    Diffuseexpand: Expanding dataset for 2d medical image segmentation using diffusion models

    Shitong Shao, Xiaohan Yuan, Zhen Huang, Ziming Qiu, Shuai Wang, and Kevin Zhou. Diffuseexpand: Expanding dataset for 2d medical image segmentation using diffusion models. arXiv preprint arXiv:2304.13416, 2023

    2023 arXiv

  31. [39]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In International Conference on Learning Representations , kigali, rwanda, 2023. Open- Review.net. 1, 4, 8, 2

    2023

  32. [40]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 8

    2023 arXiv

  33. [41]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. In International Conference on Learning Represen- tations, kigali, rwanda, 2023. OpenReview.net. 1, 4

    2023

  34. [42]

    Introducing stable diffusion 3.5

    Stability.ai. Introducing stable diffusion 3.5. https: //stability.ai/news/introducing- stable- diffusion-3-5, 2024. 7, 5

    2024

  35. [43]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063,

  36. [44]

    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. 1, 3, 4, 8, 6

    2024 arXiv

  37. [45]

    Gemma: Open models based on gemini research and tech- nology

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi `ere, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and tech- nology. arXiv preprint arXiv:2403.08295, 2024. 1

    2024 arXiv

  38. [47]

    Accelerating auto- regressive text-to-image generation with training-free spec- ulative jacobi decoding

    Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating auto- regressive text-to-image generation with training-free spec- ulative jacobi decoding. arXiv preprint arXiv:2410.01699 ,

    arXiv

  39. [48]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1

    2023 arXiv

  40. [49]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information pro- cessing systems, 30, 2017. 4

    2017

  41. [50]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023

    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, 2023. 3, 1

    2023

  42. [51]

    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 ,

    arXiv

  43. [52]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation, 2023. 3, 1

    2023

  44. [53]

    Fast sampling of dif- fusion models with exponential integrator

    Qinsheng Zhang and Yongxin Chen. Fast sampling of dif- fusion models with exponential integrator. In International Conference on Learning Representations . OpenReview.net,

  45. [54]

    Exploring self-attention for image recognition

    Hengshuang Zhao, Jiaya Jia, and Vladlen Koltun. Exploring self-attention for image recognition. In Computer Vision and Pattern Recognition, pages 10076–10085, 2020. 1

    2020

  46. [55]

    Dpm- solver-v3: Improved diffusion ode solver with empirical model statistics

    Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm- solver-v3: Improved diffusion ode solver with empirical model statistics. Advances in Neural Information Process- ing Systems, 36:55502–55542, 2023. 8 11 Bag of Design Choices for Inference of High-Resolution Masked Gener...

    2023

  47. [56]

    Benchmark and Evaluation Metrics In this section, we provide an overview of the benchmarks, evaluation metrics, and related content used in our main pa- per. 7.1. Benchmark HPD v2. The Human Preference Dataset v2 [50] is a large-scale dataset with clean annotations that focuse...

  48. [57]

    Ineffective Method Here, we summarize a collection of algorithms that demon- strate limited effectiveness when applied to MGT, aiming to provide valuable insights for other researchers. 8.1. Deterministic Sampling Deterministic sampling techniques, such as DDIM [39], have been...

  49. [58]

    Additional Information of Effective Method Here, we present discussions, analyses, and experimental results that could not be developed due to space limitations in the main paper. 9.1. Additional Experiments of Different Noise Schedule Subset of HPD v2 Method HPS v2 (↑) ImageR...

    1949

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.