Pith. sign in

REVIEW 3 major objections 6 minor 43 references

Cluster-level truncation sampling unlocks higher sample diversity in autoregressive text-to-image models while keeping quality and prompt alignment intact.

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

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

T0 review · grok-4.5

2026-07-14 10:58 UTC pith:YTJM6UU3

load-bearing objection Solid first multi-model study of sample diversity in next-token AR T2I plus a simple cluster-level p-less fix that actually moves the Pareto frontier for most of the models they tried. the 3 major comments →

arxiv 2607.10535 v1 pith:YTJM6UU3 submitted 2026-07-12 cs.CV

Improving Sample Diversity in Autoregressive Text-to-Image Generation via Cluster Truncation

Trang Nguyen , Shuang Wu , Runyan Tan , Phillip Howard This is my paper
classification cs.CV
keywords autoregressive text-to-imagesample diversitydecoding strategiescluster truncationp-less samplingvisual token codebookdiversity-quality tradeoff
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Diffusion text-to-image models often collapse sample diversity: repeated draws for one prompt look nearly the same. Autoregressive models generate images token by token and can in principle reuse language-model decoding tricks, yet the authors show those token-level tricks fail here because visual tokens stay high-entropy and are highly redundant. They introduce p-less cluster: group similar codebook vectors offline, truncate the resulting cluster probabilities by entropy, then sample a cluster first and a token inside it second. Across four AR models and two prompt suites this yields the largest diversity gains of the methods tested while quality and alignment remain acceptable. The practical payoff is a better diversity–quality frontier for creative use cases that need several distinct candidates per prompt.

Core claim

p-less cluster—a decoding strategy that partitions the visual token vocabulary into clusters of similar codes and then performs entropy-based truncation and two-stage sampling at the cluster level rather than the token level—unlocks the greatest per-prompt sample diversity across most evaluated autoregressive text-to-image models and datasets while maintaining image quality and prompt alignment.

What carries the argument

p-less cluster: offline K-means on codebook vectors, followed by aggregation of token probabilities into cluster probabilities, Rényi-entropy (p-less) truncation of those clusters, then two-stage sampling (draw a surviving cluster, then a token inside it). The mechanism turns high within-cluster visual redundancy into controlled cross-cluster exploration.

Load-bearing premise

The method assumes offline clusters of codebook vectors group tokens that are nearly interchangeable, so within-cluster swaps barely change the image while cross-cluster swaps change it a lot.

What would settle it

If, on the same models and prompt sets, p-less cluster fails to raise DINOv2 or Vendi diversity above default and dynamic-temperature baselines while VLM quality and alignment stay above 3.0, or if within-cluster token swaps produce image distances as large as cross-cluster swaps, the central claim is false.

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

If this is right

  • Autoregressive T2I models already beat the tested diffusion baselines on sample diversity; decoding can widen that gap further.
  • Complex and open-ended prompts receive the largest diversity gains from cluster-level truncation.
  • Lower classifier-free guidance plus p-less cluster can raise diversity still more when quality remains acceptable.
  • Temperature can be pushed higher under cluster truncation before quality collapses.
  • Static offline clustering of the codebook is enough; per-step re-clustering is not required for the reported gains.

Where Pith is reading between the lines

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

  • The same cluster-level truncation idea may transfer to other high-redundancy discrete spaces such as audio or video tokens where token-level nucleus sampling underperforms.
  • If future AR models use larger or more continuous-like codebooks, the optimal cluster count will likely need to scale with vocabulary size to keep the redundancy assumption valid.
  • Creative interfaces that display multiple candidates side-by-side would see immediate practical value if AR backends adopted cluster truncation by default.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies per-prompt sample diversity in next-token autoregressive text-to-image models and argues that high token-level entropy plus redundancy in visual codebooks limit standard token-level truncation methods. It proposes p-less cluster: offline K-means partitioning of the codebook, p-less truncation on the induced cluster distribution, then two-stage sampling (cluster then token). Experiments on four AR models (Anole, Emu3, Janus Pro-7B, LLaMAGen) and two prompt suites (GenEval, PartiPrompts), with diffusion baselines (Flux), report DINOv2/Vendi diversity, VLM quality/alignment, and CLIPScore, plus temperature sweeps, category breakdowns, cluster-size ablations, qualitative examples, and a short HSLP-token argument. The central claim is that p-less cluster yields the highest diversity for most model–dataset pairs while keeping quality and alignment acceptable (VLM scores typically >3).

Significance. If the results hold, this is a useful and timely contribution: the first systematic decoding study for sample diversity in token-wise AR T2I, with a simple inference-time method that exploits visual codebook geometry rather than only token probabilities. Strengths include multi-model coverage across three codebook families, bootstrap standard errors, Pareto-style temperature sweeps, prompt-category analysis, and an explicit (if brief) theoretical account of High-Support Low-Probability tokens that survive only under cluster-level truncation. The work is practically relevant for creative multi-sample interfaces and for understanding why text-style decoding does not transfer cleanly to images. It does not claim a new training paradigm or a universal fix for diversity collapse, but it does push the reported diversity–quality frontier for AR models under i.i.d. sampling.

major comments (3)
  1. [Introduction / Results] Introduction (p. 2) states that results are substantiated with human evaluations, but the manuscript never reports a human study (protocol, raters, agreement, or scores). Either add the human evaluation results with the same model/dataset coverage as the automatic metrics, or remove/qualify that claim. Automatic VLM-as-judge scores alone cannot fully substitute for the promised human evidence.
  2. [§5.1, Tables 1, 5–6] §5.1 and Tables 1, 5–6: the abstract and main claim say p-less cluster unlocks the greatest diversity across most models and datasets, yet on GenEval at default CFG it does not beat Default for Emu3 (tied DINOv2, slightly lower Vendi) and underperforms Default for Janus Pro on both diversity metrics while Dynamic Temperature collapses diversity further. The paper already notes mixed gains and stronger effects on Complex/Imagination prompts and at low CFG (Table 2), but the abstract and §5.1 framing should state more precisely when the method wins (e.g., open-ended prompts; low-CFG Janus) versus when Default remains competitive, so the central claim matches the tables.
  3. [§3.3, Figure 2(b)] §3.3 Observation 2 / Figure 2(b): the method rests on offline K-means of static codebook vectors producing perceptually redundant clusters. Figure 2(b) is the main empirical support (within-cluster swaps keep DINOv2 distance <0.4). That evidence is for Janus Pro only and uses random swaps rather than sampling under the actual truncated distribution. A short multi-model check (or at least one additional codebook family) that within-cluster vs cross-cluster swaps behave similarly would make the load-bearing geometric assumption more robust; otherwise the generality of cluster-level truncation remains partly model-specific.
minor comments (6)
  1. [§4.3, Appendix E] §4.3 and Appendix E: VLM-as-judge (InternVL3-14B, 1–5) is well motivated given CLIPScore’s weakness on artifacts, but inter-run variance and correlation with human preference are not quantified beyond averaging three runs. A brief reliability note would help.
  2. [Appendix C, Table 4] Table 4 / Appendix C: optimal (T, k, n) differ by model and dataset; the 2×2 grid is small and practical, but the main text should state more clearly that reported “best” diversity points are after this per-setting selection on a held-out prompt subset, not a single fixed hyperparameter set.
  3. [Figure 4, Figure 11] Figure 4 and Figure 11: Pareto plots use a reduced prompt subset (120/90). State that explicitly in the figure captions so readers do not equate them with full-benchmark Tables 1–2.
  4. [Appendix D / C] Typo: “Addtional experimental results” (Appendix D heading); “Hyperpameter” (C.2); “Partiprompts” inconsistently capitalized. Clean these up.
  5. [§2.1–2.2] Related work §2.1 correctly notes DiverseAR targets bitwise/next-scale AR; a one-sentence note on whether cluster truncation could transfer to next-scale or mask-based AR would clarify scope without expanding experiments.
  6. [Appendix H] Appendix H Proposition H.1 uses Rényi-2 collision probability as the p-less threshold; the main method allows k∈{1.5,2.0}. Align the proposition statement with the k used in experiments or note that the HSLP argument is for the k=2 case.

Circularity Check

0 steps flagged

No significant circularity: empirical decoding method with comparative tables, not a first-principles derivation that reduces to its inputs.

full rationale

This is an empirical methods paper proposing a cluster-level truncation sampler (p-less cluster) for autoregressive T2I models. The central claim is comparative performance (higher DINOv2/Vendi diversity while VLM quality/alignment remain acceptable) across four models and two datasets, not a prediction derived from fitted parameters or a uniqueness theorem. Hyperparameters (k, n, T) are tuned on a held-out prompt subset and then evaluated on the full benchmarks; that is ordinary model selection, not a fitted input renamed as prediction. Observation 2 (within-cluster token swaps produce small image change; cross-cluster swaps produce large change) is an empirical measurement on the codebook (Figure 2(b)), not a definition that forces the diversity gains. Appendix H's HSLP proposition is a short inequality proof under stated conditions (HSLP-1, HSLP-2) showing that certain tokens receive strictly higher probability under cluster-level p-less than under token-level p-less; the proof does not assume the experimental outcome. Self-citations to the authors' prior p-less work supply the truncation rule but are not load-bearing uniqueness claims that forbid alternatives. No step reduces by construction to its own inputs. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on standard ML evaluation practice plus two modeling choices (offline K-means geometry equals perceptual redundancy; VLM scores are reliable proxies) and a small set of free hyper-parameters that are tuned per model. No new physical entities are postulated; the only invented construct is the HSLP token class used in the appendix proof.

free parameters (3)
  • number of clusters n = typically 500
    Chosen from {500,1000} (sometimes extended to 2000) per model/dataset to maximize diversity under a quality gate; directly controls truncation aggressiveness.
  • Rényi order k = 1.5 or 2.0
    Chosen from {1.5,2.0}; sets the entropy threshold for cluster truncation.
  • sampling temperature T = model-specific, often 2.5–4.0
    Swept and selected jointly with n,k so that VLM quality/alignment stay above 3.0 while diversity is maximized.
axioms (3)
  • domain assumption Offline K-means on codebook vectors yields clusters of visually redundant tokens
    Stated as Observation 2 and validated only by a small swap experiment (Figure 2b); load-bearing for why cluster-level truncation should increase image diversity.
  • domain assumption VLM-as-judge (InternVL3-14B) scores on a 1–5 scale are reliable proxies for human perceptual quality and prompt alignment
    Used as the primary quality/alignment gate; paper notes CLIPScore limitations but does not validate the VLM scores against humans beyond claiming they are more discriminative.
  • domain assumption i.i.d. sampling without cross-sample coupling is the correct setting for measuring sample diversity
    Explicit design choice in §3.1; excludes methods that optimize a set jointly.
invented entities (2)
  • p-less cluster decoding no independent evidence
    purpose: Two-stage truncation-and-sampling procedure that first filters clusters then samples a token inside the chosen cluster
    The core algorithmic contribution; defined in §3.2.
  • High-Support Low-Probability (HSLP) tokens no independent evidence
    purpose: Tokens filtered by token-level p-less but retained by cluster-level p-less; used to prove a strict probability inequality
    Introduced in Appendix H to justify why cluster-level truncation can recover diversity.

pith-pipeline@v1.1.0-grok45 · 24309 in / 2774 out tokens · 32519 ms · 2026-07-14T10:58:43.816082+00:00 · methodology

0 comments
read the original abstract

While diffusion models achieve state-of-the-art image quality for text-to-image (T2I) generation, recent work has demonstrated that they suffer from sample diversity collapse. In this work, we investigate whether autoregressive (AR) image generation models can push the Pareto frontier between image quality and sample diversity. With recent advances in quality and efficiency, AR models have emerged as a viable alternative to diffusion-based image generation. Beyond enabling new use cases such as interleaved image-text generation, their sequential generation process makes them compatible with a wide range of token-based decoding strategies originally developed to improve diversity in text generation. Motivated by the potential of a better diversity-quality tradeoff in the AR paradigm, we present the first systematic study of sample diversity in AR image generation models. We show that two key properties of AR image generation, persistently high token-level entropy and substantial redundancy in visual token spaces, limit the effectiveness of existing token-level decoding methods for diversity enhancement. We therefore propose $p$-less cluster, a new decoding strategy that performs entropy-based truncation sampling at cluster level rather than at token level. We evaluate our approach and baseline decoding methods across four autoregressive T2I models and two datasets using a comprehensive suite of metrics spanning image quality, prompt alignment, and diversity. Our results show that $p$-less cluster unlocks the greatest diversity across most evaluated autoregressive T2I models and datasets while maintaining image quality and prompt alignment.

Figures

Figures reproduced from arXiv: 2607.10535 by Phillip Howard, Runyan Tan, Shuang Wu, Trang Nguyen.

Figure 1
Figure 1. Figure 1: p-less cluster enables diverse i.i.d. sampling across autoregressive (AR) image generative models while maintaining acceptable image quality, avoiding the issue of diversity collapse. generation: by formulating image synthesis as sequential token prediction, AR models can leverage decoding strategies developed previously for language generation to promote diversity. In the realm of text generation, a wide … view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of decoding strategies on image- and token-level diversity (Janus Pro). (a) Visual comparison; (b) impact of within- vs. cross-cluster token swaps on image diversity; (c) percentage of distinct tokens across generation steps (100 samples). At high temperature, default sampling admits too many tokens, leading to degeneration. Dynamic temperature mitigates this by reducing the effective temperatur… view at source ↗
Figure 3
Figure 3. Figure 3: DINOv2 diversity by PartiPrompts challenge category. p-less cluster achieves larger gains in the Complex and Imagination categories with more open-ended and abstract prompts [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Diversity–alignment tradeoff across decoding strategies for Anole and Janus Pro using GenEval Dataset (DINOv2 vs. Prompt alignment). Janus Pro-7B uses low CFG due to its low diversity under default CFG (Section 5.1). p-less cluster shifts the Pareto frontier toward higher diversity while remaining competitive on alignment while Dynamic Temperature maintains same level of high alignment but low diversity ac… view at source ↗
Figure 5
Figure 5. Figure 5: Effect of number of clusters on prompt alignment (left) and diversity (right) for Anole on PartiPrompts. Smaller number of clusters increases robustness with respect to image quality at higher temperatures and exhibits more stability in sample diversity. To investigate the effect of cluster size, we fix k = 2.0 and conduct experiments on an extended range of cluster size {500, 1000, 1500, 2000}. As illustr… view at source ↗
Figure 6
Figure 6. Figure 6: Shannon entropy distribution over decoding steps for text vs. image generation in Janus Pro, aggregated across 10 prompts. The image generation mode sustains substantially higher per-token entropy throughout decoding compared to the text generation mode under identical temperature and CFG settings, confirming that AR image models operate in a fundamentally higher￾entropy regime [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative and quantitative comparison of how different decoding strategies affect image-level and token-level diversity (Janus Pro). At high temperature, both Default and Default cluster sampling admits too many tokens which leads to degeneration. Both Dynamic Temperature and p-less mitigates quality issues at the cost of limited diversity due to low number of distinct tokens. p-less cluster strikes a be… view at source ↗
Figure 8
Figure 8. Figure 8: Robustness of decoding strategies across sampling temperatures with respect to Diversity (top) and Alignment (bottom) on GenEval. Starred markers indicate the highest-diversity point that remains above the alignment gate (dashed line at VLM Alignment > 3.0). p-less cluster (green) remains above the quality gate across a wider range of temperatures than Default sampling, though narrower than Dynamic Tempera… view at source ↗
Figure 9
Figure 9. Figure 9: Robustness of decoding strategies across sampling temperatures with respect to Diversity (top) and Alignment (bottom) on Partiprompts. Starred markers indicate the highest￾diversity point that remains above the alignment gate (dashed line at VLM Alignment > 3.0). p-less cluster (green) remains above the quality gate across a wider range of temperatures than Default sampling, though narrower than Dynamic Te… view at source ↗
Figure 10
Figure 10. Figure 10: DINOv2 diversity by PartiPrompts challenge category across all four AR models. p-less cluster consistently achieves the highest diversity in the Complex and Imagination categories across all models. For Anole and LlamaGen, the gains extend uniformly across all categories, whereas for Emu3 and Janus Pro the improvements are more concentrated in open-ended prompts [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Diversity–alignment tradeoff across decoding strategies for all 4 models using GenEval Dataset (DINOv2 vs. Prompt alignment) . Janus Pro-7B uses low CFG (CFG=1.0) due to its low diversity under default CFG (Section 5.1). Our p-less cluster method (triangles) shifts the Pareto frontier toward higher diversity while remaining competitive on alignment (except for Emu3). 19 [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison for Emu3. Hyperparameters match [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison for Anole. Top: GenEval prompts; Bottom: PartiPrompts. Hyperparameters match [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison for Janus at default CFG. Top: GenEval prompts; Bottom: PartiPrompts. Hyperparameters match [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison for LLaMAGen. Top: GenEval prompts; Bottom: PartiPrompts. Hyperparameters match [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Fraction of tokens satisfying the HSLP conditions ((HSLP-1), (HSLP-2)) for Janus Pro-7B across decoding steps. A reasonable portion of tokens (approximately 10%) satisfy both conditions, indicating that they are filtered by token-level p-less but would survive cluster-level p-less, thus receiving a non-zero probability under our method. H.1 Empirical prevalence of HSLP tokens [PITH_FULL_IMAGE:figures/ful… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 14 linked inside Pith

  1. [1]

    S. AI. Stable diffusion 3.5 technical release. https://stability.ai/news/ introducing-stable-diffusion-3-5, 2024. Accessed: 2026-05-06

    2024

  2. [2]

    Angelova

    N. Angelova. Integration of artificial intelligence into traditional graphic designer software products–a relief and a challenge. InAIP Conference Proceedings, volume 3274, page 080008. AIP Publishing LLC, 2025

    2025

  3. [3]

    X. Chen, Z. Wu, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, and C. Ruan. Janus-pro: Uni- fied multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811, 2025. URLhttps://arxiv.org/abs/2501.17811

    Pith/arXiv arXiv 2025

  4. [4]

    Chern, Z

    E. Chern, Z. Su, et al. Anole: An open, autoregressive, native large multimodal models for interleaved image-text generation.arXiv preprint arXiv:2407.06135, 2024. URL https: //arxiv.org/abs/2407.06135

    Pith/arXiv arXiv 2024

  5. [5]

    J. Cho, Y . Hu, R. Garg, P. Anderson, R. Krishna, J. Baldridge, M. Bansal, J. Pont-Tuset, and S. Wang. Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation.arXiv preprint arXiv:2310.18235, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    A. Fan, M. Lewis, and Y . Dauphin. Hierarchical neural story generation.arXiv preprint arXiv:1805.04833, 2018

    Pith/arXiv arXiv 2018

  7. [7]

    Friedman and A

    D. Friedman and A. B. Dieng. The vendi score: A diversity evaluation metric for machine learning.arXiv preprint arXiv:2210.02410, 2022

    Pith/arXiv arXiv 2022

  8. [8]

    Gandikota and D

    R. Gandikota and D. Bau. Distilling diversity and control in diffusion models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1304–1313, March 2026

    2026

  9. [9]

    Ghosh et al

    A. Ghosh et al. Geneval: An object-focused framework for evaluating text-to-image alignment. InAdvances in Neural Information Processing Systems, 2023. URL https://arxiv.org/ abs/2310.11513

    Pith/arXiv arXiv 2023

  10. [10]

    J. Han, J. Liu, J. Yi, B. Yan, Y . Zhang, Z. Yuan, B. Peng, and X. Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15733– 15744, 2025. URL https://openaccess.thecvf.com/content/CVPR2025/html/Han_ Infinity_Scaling_Bi...

    2025

  11. [11]

    Y . He, F. Chen, Y . He, S. He, H. Zhou, K. Zhang, and B. Zhuang. Zipar: Accelerating auto-regressive image generation through spatial locality.arXiv preprint arXiv:2412.04062, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InAdvances in Neural Information Processing Systems, volume 30, pages 6626–6637, 2017

    2017

  13. [13]

    Holtzman, J

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y . Choi. The curious case of neural text degeneration. InInternational Conference on Learning Representations

  14. [14]

    Y . Hu, B. Liu, J. Kasai, Y . Wang, M. Ostendorf, R. Krishna, and N. A. Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406–20417, 2023

    2023

  15. [15]

    B. F. Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  16. [16]

    T. Li, Y . Tian, H. Li, M. Deng, and K. He. Autoregressive image generation without vector quantization.Advances in Neural Information Processing Systems, 37:56424–56445, 2024

    2024

  17. [17]

    J. Liu, J. Han, B. Yan, F. Zhu, X. Wang, Y . Jiang, B. PENG, Z. Yuan, et al. Infinitystar: Unified spacetime autoregressive modeling for visual generation.Advances in neural information processing systems, 38, 2025. 10

    2025

  18. [18]

    X. Ma, F. Zhao, P. Ling, H. Qiu, Z. Wei, H. Yu, J. Huang, Z. Zeng, and L. Ma. Towards better & faster autoregressive image generation: From the perspective of entropy. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URL https: //openreview.net/forum?id=LiQH1MOCMs

    2026

  19. [19]

    J. Mu, N. Vasconcelos, and X. Wang. Editar: Unified conditional generation with autoregressive models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7899–7909, 2025

    2025

  20. [20]

    M. N. Nguyen, A. Baker, C. Neo, A. Roush, A. Kirsch, and R. Shwartz-Ziv. Turning up the heat: Min-p sampling for creative and coherent llm outputs.arXiv preprint arXiv:2407.01082, 2024

    arXiv 2024

  21. [21]

    Oquab, T

    M. Oquab, T. Darcet, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023. URLhttps://arxiv.org/abs/2304.07193

    Pith/arXiv arXiv 2023

  22. [22]

    Parmar, O

    G. Parmar, O. Patashnik, D. Ostashev, K.-C. J. Wang, K. Aberman, S. Narasimhan, and J.-Y . Zhu. Scaling group inference for diverse and high-quality generation. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview. net/forum?id=IyTNxjTuWT

    2026

  23. [23]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    Pith/arXiv arXiv 2022

  24. [24]

    Razavi, A

    A. Razavi, A. Van den Oord, and O. Vinyals. Generating diverse high-fidelity images with vq-vae-2.Advances in neural information processing systems, 32, 2019

    2019

  25. [25]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gon- tijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35: 36479–36494, 2022

    2022

  26. [26]

    P. Sun, Y . Jiang, Z. Chen, J. Yuan, et al. Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 2024. URL https://arxiv. org/abs/2406.06525

    Pith/arXiv arXiv 2024

  27. [27]

    R. Tan, S. Wu, and P. Howard. p-less sampling: A robust hyperparameter-free approach for llm decoding. InThe Fourteenth International Conference on Learning Representations

  28. [28]

    R. Tan, S. Wu, and P. Howard. p-less sampling: A robust hyperparameter-free approach for llm decoding.arXiv preprint arXiv:2509.23234, 2025. URL https://arxiv.org/abs/2509. 23234. [cs.AI, cs.CL]

    arXiv 2025

  29. [29]

    Y . Teng, F. Wang, X. Liu, et al. Speculative jacobi-denoising decoding for accelerating autore- gressive text-to-image generation. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=y2eWc6jrlu

    2025

  30. [30]

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems, 37:84839–84865, 2024

    2024

  31. [31]

    Van den Oord, N

    A. Van den Oord, N. Kalchbrenner, L. Espeholt, O. Vinyals, A. Graves, et al. Conditional image generation with pixelcnn decoders.Advances in neural information processing systems, 29, 2016

    2016

  32. [32]

    Van Den Oord, N

    A. Van Den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747–1756. PMLR, 2016

    2016

  33. [33]

    F. Wang, Y . Teng, X. Liu, et al. Diverse text-to-image generation via contrastive noise opti- mization. InThirty-ninth Conference on Neural Information Processing Systems, 2025. URL https://openreview.net/forum?id=EVRMnAREc3

    2025

  34. [34]

    X. Wang, Z. Zhang, et al. Emu3: Next-token prediction is all you need.arXiv preprint arXiv:2409.18869, 2024. URLhttps://arxiv.org/abs/2409.18869. 11

    Pith/arXiv arXiv 2024

  35. [35]

    Y . Yang, Z. Lv, T. Pan, H. Wang, B. Yang, H. Yin, C. Li, and C. Si. Diversear: Boosting diversity in bitwise autoregressive image generation.arXiv preprint arXiv:2512.02931, 2025

    arXiv 2025

  36. [36]

    T. Yao, Y . Li, Y . Pan, Z. Qiu, and T. Mei. Denoising token prediction in masked autoregressive models. InProceedings of the IEEE/CVF international conference on computer vision, pages 18024–18033, 2025

    2025

  37. [37]

    T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman. Improved distribution matching distillation for fast image synthesis. InNeurIPS, 2024

    2024

  38. [38]

    J. Yu, X. Li, J. Y . Koh, H. Zhang, R. Pang, J. Qin, A. Ku, Y . Xu, J. Baldridge, and Y . Wu. Vector- quantized image modeling with improved vqgan. InInternational Conference on Learning Representations, 2022

    2022

  39. [39]

    J. Yu, Y . Xu, J. Y . Koh, T. Luong, G. Baid, et al. Scaling autoregressive models for content- rich text-to-image generation. InTransactions on Machine Learning Research, 2022. URL https://arxiv.org/abs/2206.10789

    Pith/arXiv arXiv 2022

  40. [40]

    Zhang, Y

    S. Zhang, Y . Bao, and S. Huang. Edt: Improving large language models’ generation by entropy-based dynamic temperature sampling (2024).URL https://arxiv. org/abs/2403.14541

    Pith/arXiv arXiv 2024

  41. [41]

    Topic: {prompt}. Generate a descriptive image caption using N tokens

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y . Duan, W. Su, J. Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025. 12 A Limitations & Broader Impacts Limitations.Our approach operates purely at inference time, and its effectiveness is therefore bou...

    Pith/arXiv arXiv 2025

  42. [42]

    Visual quality (1–5) Assess the intrinsic quality of the image, independent of the prompt: •5 – Excellent: Sharp, well-lit, natural colors, zero artifacts, professional-grade composition •4 – Good: Minor imperfections (slight blur, minor noise) that don’t detract from the image •3 – Acceptable: Noticeable issues (soft focus, unnatural lighting, minor dist...

  43. [43]

    quality_score

    Prompt alignment (1–5) Assess how faithfully the image represents the prompt. Consider: subject, attributes (color, size, texture), spatial relationships, style, and mood: •5 – Perfect: All prompt elements present and accurately depicted •4 – Good: Core elements present; minor attributes missing or slightly off •3 – Partial: Main subject recognizable but ...