Pith. sign in

REVIEW 4 major objections 5 minor 47 references

In scale-wise autoregressive image models, guidance drifts away from salient tokens; an attention-weighted reweighting of guidance consistently beats classifier-free guidance.

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

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

T0 review · deepseek-v4-flash

2026-08-04 14:38 UTC pith:CSEWIPDC

load-bearing objection A plausible training-free guidance reweighting for scale-wise AR models with a genuinely useful diagnostic lens, but the headline gain over CFG is currently confounded by unmatched guidance scales and an underspecified attention reduction. the 4 major comments →

arxiv 2509.23876 v3 pith:CSEWIPDC submitted 2025-09-28 cs.CV cs.AI

Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance

Ky Dan Nguyen , Hoang Lam Tran , Anh-Dung Dinh , Daochang Liu , Weidong Cai , Xiuying Wang , Chang Xu This is my paper
classification cs.CV cs.AI
keywords visual autoregressive modelsscale-wise autoregressionclassifier-free guidanceinformation-grounding guidanceattention weightingimage generationtext-to-imageguidance misalignment
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.

This paper tries to show that classifier-free guidance, the standard way to condition autoregressive image models, silently degrades as sampling proceeds: guidance signals aimed at all tokens equally become dispersed and drift away from the foreground objects they should describe. The authors propose Information-Grounding Guidance (IGG), which reweights the guidance given to each token with a self-attention score computed on the guidance signal itself, so that nudges concentrate where semantic content lives. Across ImageNet class-conditioned generation with the VAR backbone and text-to-image generation with VAR-CLIP and Switti, IGG reports consistent improvements over classifier-free guidance in FID, IS, and several text-alignment metrics. If the claim holds, it means that in scale-wise autoregressive models the right place to apply guidance is not the whole token map uniformly but a contextually computed subset of tokens.

Core claim

The paper's central claim is that guidance in scale-wise autoregressive (SwAR) image models is structurally misaligned: as the model predicts progressively finer token maps, classifier-free guidance signals become more even across tokens and weaker on the tokens that actually carry the object, unlike diffusion models where guidance stays sharp and object-focused. From this it concludes that not all tokens should be guided equally. The proposed remedy, Information-Grounding Guidance (IGG), writes guided prediction as the unconditional prediction plus a token-wise weighting of the guidance signal, with weights produced by self-attention over the guidance map itself. On VAR at 256x256 and 512x5

What carries the argument

The central mechanism is the guidance-reweighting equation p~_theta(s_k|c) = p_theta(s_k) + f_k(s_k|c) * p_theta^->(s_k|c), where p_theta^-> is the CFG 'nudge' between conditional and unconditional predictions. The per-token weight f_k is a softmax self-attention over the guidance signal itself, f_k = softmax( p-> (p->)^T / sqrt(|V|) ), which assigns higher weight to tokens whose guidance pattern resembles that of other salient tokens. This same attention output is what concentrates guidance on foreground regions. Two diagnostic quantities — Pielou evenness and Jensen-Shannon divergence between guided and background-token distributions — carry the analysis by quantifying how much guidance fo

Load-bearing premise

The central claim presupposes that the quality gains come from concentrating guidance on semantically important tokens; the experiments compare against CFG at different guidance strengths, so the attention mechanism may or may not be the source of the improvement.

What would settle it

Run classifier-free guidance on VAR-d36 with w=2.10 (IGG's scale) and on VAR-d30 with w=1.85, then compare FID and IS against IGG at the same weights. If CFG at those scales matches IGG, the reweighting mechanism is not the cause; alternatively, sweep the guidance scale for both methods and check whether IGG's FID-IS curve dominates CFG's at matched diversity (IS) rather than at each method's separate optimum.

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

If this is right

  • IGG can be dropped into any SwAR model (VAR, VAR-CLIP, Switti) at inference time without retraining, since it only modifies how the existing conditional and unconditional logits are combined.
  • Because IGG improves more at 512x512 than at 256x256, the benefit should grow for higher-resolution SwAR sampling where token maps are larger.
  • The evenness-divergence equilibrium identifies near-optimal guidance weights (w* ~ 1.35 for VAR-d30-IGG, giving FID ~1.98), suggesting a cheap tuning rule that avoids full FID sweeps.
  • In text-to-image generation, IGG raises GenEval on Switti (0.62 to 0.64) and lowers FID, implying better prompt-following and fewer text-conditioned artefacts.
  • Mixing CFG with a negative IGG weight slightly improves diffusion models (EDM2-S FID 2.29 to 2.20), showing the reweighting idea transfers outside SwAR.

Where Pith is reading between the lines

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

  • If attention reweighting—rather than the higher guidance scale—is what matters, then other cheap saliency signals (segmentation maps, gradient norms) that identify foreground tokens should reproduce part of IGG's gain; this is testable without attention.
  • The paper's own diagnostic pair, evenness and divergence, could be turned into a no-reference, tuning-free guidance selector: pick the weight where the two scaled scores cross, rather than searching FID.
  • The insight that guidance should be spatially concentrated suggests that for SwAR text-to-image, conditioning tokens (such as CLIP text features) could be used to compute f_k instead of self-attention on the guidance map, potentially making the method cheaper and more prompt-aligned.
  • One open question the paper leaves is why a negative IGG weight helps diffusion; if the denoising-acceleration explanation is right, the sign of the reweighting should flip with the noise level.

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

4 major / 5 minor

Summary. The paper studies classifier-free guidance (CFG) in scale-wise autoregressive (SwAR) image generation. It argues that, unlike in diffusion models, CFG guidance signals in SwAR models are dispersed and progressively weaken, and that applying a single global guidance scale to all tokens is suboptimal. To fix this, the authors propose Information-Grounding Guidance (IGG), which reweights the CFG nudge at each token by a self-attention-derived importance mask (Eq. 5-6). They introduce two diagnostic metrics, Pielou evenness and Jensen-Shannon divergence, to quantify guidance concentration. IGG is evaluated on class-conditioned ImageNet with VAR backbones and on text-to-image benchmarks with VAR-CLIP and Switti, reporting improved FID/IS and prompt-following over CFG. The paper also includes ablations on guidance schedules, sliding-window attention, and a mixed CFG-IGG variant for diffusion models.

Significance. If the central claim holds, IGG would be a simple, training-free, attention-based guidance mechanism that is the first designed specifically for SwAR models, with a plausible mechanistic story and publicly released code. The diagnostic evenness/divergence metrics, if properly defined, could be useful for monitoring guidance behavior and even for choosing guidance scales. The paper is careful to use off-the-shelf pretrained models and reports a broad set of metrics. However, the headline improvements rest on a comparison that is confounded by different guidance scales, and the formulation of the core IGG mechanism contains a dimensional gap. These issues must be resolved before the empirical claims and the mechanism can be accepted.

major comments (4)
  1. [§6.1, Table 2] The central claim that IGG outperforms CFG on VAR is not supported by the current comparison because the two methods are evaluated at different guidance scales. For VAR-d30, CFG uses w=1.75 while IGG uses w=1.85; for VAR-d36, CFG uses w=1.50 while IGG uses w=2.10. The FID differences are small (0.01 and 0.05, respectively), and §6.3 shows that for VAR-d30-IGG the FID varies from ≈1.98 at w=1.35 to ≈1.92 at w=1.85, a swing larger than the reported CFG-vs-IGG gap. Without evaluating CFG at the same guidance scales (or demonstrating via a full scale sweep that the advantage is not due to scale alone), the improvement cannot be attributed to the attention-based reweighting. This is load-bearing for the paper's main conclusion.
  2. [§5, Eqs. (5)-(6)] The definition of IGG is underspecified. Eq. (5) requires f_k to be a per-token weight vector in R^{h_k w_k}, but Eq. (6) computes a softmax attention matrix of shape (h_k w_k) × (h_k w_k) from the guidance signal p→_θ(s_k|c), whose last dimension is the vocabulary size |V|. The paper never states how this matrix is reduced to a per-token weight vector, nor the softmax axis, nor how the vocabulary dimension is marginalized/averaged before forming the attention matrix. Since f_k is the entire mechanism of IGG, this omission makes the method irreproducible as written. Please provide the exact tensor shapes and reduction used (e.g., row-sum, diagonal, or mean over a specific axis).
  3. [§4, Eqs. (3)-(4)] The diagnostic metrics used to motivate the method are not well defined for the objects they are applied to. In Eq. (3), PEI is the normalized Shannon entropy of the 'guidance distribution', but p→_θ(s_k|c) as defined in Eq. (2) is a signed difference of probabilities (or logits), not a nonnegative probability distribution. Likewise, Eq. (4) uses Jensen-Shannon divergence between p→ and q→, which requires probability vectors. Without specifying how p→ is normalized, clipped, or otherwise converted into a distribution, the claims that PEI∈[0,1] and JSD∈[0,1], and hence the quantitative comparisons in Table 1 and Figure 4, are not formally grounded. This affects the validity of the motivating analysis, though it is separable from the IGG mechanism itself.
  4. [§6.2, Table 3] The text-to-image results are presented without reporting the guidance scales used for CFG and IGG, and the reported differences are often within noise (e.g., COCO FID 10.95 vs 10.93; CLIP 0.264 vs 0.264). The claimed significant improvement on GenEval and FID for Switti needs the guidance schedules, number of samples, and ideally confidence intervals or multiple seeds. In its current form, this table does not provide enough evidence for the text-to-image component of the central claim.
minor comments (5)
  1. [§3, Eq. (1)] Eq. (1) uses λ_k for the guidance schedule, but the sentence below says 'where w∈R is the guidance scale'. For clarity, explicitly state the relationship λ_k = w·k/(K-1) and consistently distinguish the schedule λ_k from the scalar hyperparameter w reported in Table 2.
  2. [§5, Eq. (6)] The notation p→[p→]^T / sqrt(|V|) is ambiguous for a tensor-valued p→. Please specify the exact contraction (e.g., over the vocabulary axis) and whether the softmax is row-wise or column-wise.
  3. [§6.3, Figure 4] The 'scaled' and 'raw' evenness/divergence scores are not defined precisely. It would be helpful to state the exact scaling factor and to motivate why reciprocal scaling removes the dependence on w.
  4. [Appendix A.5, Eq. (7)] The mixture guidance scheme introduces (w, w') but does not define the schedules {γ_k, γ'_k} beyond saying they are fixed. Since Table 5 reports only w and w', please make the conversion explicit.
  5. [Throughout] Minor typos and formatting issues: 'SwAR models ... leaves more room' should be 'leave'; 'guidance signals often weakens' should be 'weaken'; reference names contain inconsistent accented characters (Voronov vs Voronov).

Circularity Check

0 steps flagged

No significant circularity: IGG is a parameter-free reweighting evaluated on external FID/IS; diagnostic metrics are descriptive and not load-bearing.

full rationale

The paper's central claim is that Information-Grounding Guidance (IGG) improves sampling quality in scale-wise autoregressive models. IGG is defined by Eq. 5 as CFG with a per-token weight f_k, and Eq. 6 computes f_k via a parameter-free self-attention operation over the guidance signal itself. No parameter is fitted to the evaluation benchmarks, and the method is not derived from the metrics it later reports. The headline comparisons against CFG use external metrics (FID, IS, CLIP, GenEval) that are independent of the proposed formulation. The evenness and divergence metrics in Section 4 are descriptive diagnostics; although they are co-designed with the method and used to illustrate the motivating hypothesis, they are not the basis of the SOTA claim. The paper does not present a mathematical derivation that reduces to its inputs, and the cited prior work (including the authors' own guidance methods) is not load-bearing for the central result. The different guidance scales used for CFG and IGG in Table 2 are a potential experimental confound, but that is a comparison-validity concern, not circularity under the definitions provided. Therefore the derivation is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 2 invented entities

The central method introduces no learned parameters, which is a strength. The main free parameters are the guidance scales, which are tuned per method and per model and differ between IGG and CFG baselines, threatening the attribution of the gains. The axioms include underspecified mathematical interpretations of p->theta and an unvalidated equivalence between segmentation masks and semantic importance.

free parameters (2)
  • guidance scale w for IGG = 1.85 (VAR-d30), 2.10 (VAR-d36); undisclosed for Switti/VAR-CLIP in Table 3
    The guidance schedule is lambda_k = w*k/(K-1). Different w values are hand-tuned for IGG and CFG separately, and the comparison in Table 2 uses stronger w for IGG. The Switti text-to-image experiments do not report w at all.
  • mixed guidance weights (w, w') for EDM2 = (1.40, -0.40) for EDM2-S; (1.20, -0.20) for EDM2-XXL
    In Appendix A.5 the CFG-IGG mixture requires an additional guidance weight w', which is hand-tuned and reported only for this diffusion experiment.
axioms (5)
  • domain assumption Attention over guidance signals identifies semantically important tokens
    Section 5 motivates f_k via the intuition that 'a token surrounded by other salient tokens should itself be considered important', and assumes the self-attention of p->theta captures salience. No proof or external validation is given.
  • domain assumption The evenness and divergence metrics are meaningful proxies for guidance quality
    Section 4 defines PEI and JSD as quality indicators and interprets a more uneven, more divergent guidance distribution as better, but this relationship is asserted rather than derived.
  • ad hoc to paper p->theta(s_k|c) can be treated as a distribution over token positions
    Eqs. 3-4 compute Shannon entropy and JSD over tokens, but p->theta is a per-token distribution over the vocabulary; the collapse from vocabulary to a single scalar per position is never defined.
  • domain assumption YOLOv11 segmentation identifies semantically important tokens
    Appendix A.1 uses YOLOv11 masks to split foreground and background tokens, implicitly equating semantic importance with segmentation masks.
  • standard math The CFG nudge interpretation in Eq. 2 is valid
    Expanding CFG as p_theta(s_k) + gamma_k (p_theta(s_k|c) - p_theta(s_k)) is algebraically equivalent to the standard interpolation, so this assumption is standard.
invented entities (2)
  • Per-token guidance weighting function f_k no independent evidence
    purpose: Reweights the CFG nudge so that semantically important tokens receive stronger guidance during sampling.
    f_k is a constructed function (Eq. 6) with no external falsifiable handle. Its exact implementation is ambiguous because the equation defines a matrix, not the required vector of weights.
  • Guidance misalignment as a causal mechanism no independent evidence
    purpose: Explains why SwAR guidance is dispersed and motivates the IGG correction.
    The paper posits that progressive resolution scaling creates information inconsistencies that scatter guidance, but this mechanism is inferred from the same metrics used to design the method, not independently measured.

pith-pipeline@v1.3.0-alltime-deepseek · 14662 in / 13631 out tokens · 334943 ms · 2026-08-04T14:38:07.574006+00:00 · methodology

0 comments
read the original abstract

Autoregressive (AR) models based on next-scale prediction have emerged as a powerful tool for image generation, but they face a critical weakness: information inconsistencies between patches across timesteps introduced by progressive resolution scaling. These inconsistencies scatter guidance signals, causing them to drift away from salient regions within the image and leaving behind ambiguous, unfaithful features during sampling. We tackle this challenge with Information-Grounding Guidance (IGG), a novel framework that anchors guidance to semantically important tokens via an attention-based dynamic weighting formulation, consequently ensuring that guidance and semantic contents remain tightly aligned. Across both class-conditioned and text-to-image generation tasks, IGG delivers sharper, more coherent, and semantically grounded images, demonstrating its efficacy for correcting AR-based methods. Our code is available at https://github.com/dnngky/infoground-guidance.

Figures

Figures reproduced from arXiv: 2509.23876 by Anh-Dung Dinh, Chang Xu, Daochang Liu, Hoang Lam Tran, Ky Dan Nguyen, Weidong Cai, Xiuying Wang.

Figure 1
Figure 1. Figure 1: Comparison between classifier-free guidance (top) and our method (bottom) on ImageNet 512 × 512 class-conditioned generation (class: cock), with VAR (Tian et al., 2024) as the backbone model. Each column corresponds to a sampling step. Each heat map depicts the distribution of guidance on tokens at the respective sampling step, ranging from purple (weak guidance) to yellow (strong guidance). Blue and red s… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of guidance throughout the sampling process of EDM2 (Karras et al., 2024b) (top) and VAR (Tian et al., 2024) (bottom) on ImageNet 512 × 512 class-conditioned generation (class: sports car). For the sake of comparison, the original number of sampling steps of EDM2 (32 steps) has been modified to match VAR. Sampling steps are respectively labelled with their evenness and divergence scores, with … view at source ↗
Figure 4
Figure 4. Figure 4: Analysing the relationships between various metric scores attained by VAR-d30-IGG. From left to right: effect of changing guidance scales on reported evenness and divergence scores, where dashed and solid lines depict raw and scaled scores respectively; correspondence between FID and scaled evenness and divergence scores; and FID-IS trade-off curve. Guidance scale vs. evenness and divergence. To better und… view at source ↗
Figure 3
Figure 3. Figure 3: Example 1024 × 1024 generations of Switti under three guidance schemes: no guidance, CFG, and IGG (ours). Without IGG, CFG or vanilla sampling of Switti has higher chance of generating failure features, around one in four samples. Evenness and divergence vs. FID. We also investigated the relationship between the proposed metrics and FID, as depicted in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional comparisons of guidance signals in EDM2 (Karras et al., 2024b) (top) and VAR (Tian et al., 2024) (bottom) [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional comparisons of sampling in VAR-d36 (Tian et al., 2024) using CFG and IGG. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualisation of guidance signals for every timestep in the original sampling pro￾cess of EDM2 (Karras et al., 2024b). Timesteps falling within the guidance interval reported in Kynka¨anniemi et al. (2024) are highlighted in orange. This interval coincides with the transition ¨ from noise to image, which is where guidance plays the most influential role. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] 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

47 extracted references · 1 canonical work pages

  1. [1]

    Peters, and Arman Cohan

    Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The Long - Document Transformer , December 2020. URL http://arxiv.org/abs/2004.05150. arXiv:2004.05150 [cs]

    Pith/arXiv arXiv 2020

  2. [3]

    Generative Pretraining From Pixels

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative Pretraining From Pixels . In Proceedings of the 37th International Conference on Machine Learning , pp.\ 1691--1703. PMLR, 2020. URL https://proceedings.mlr.press/v119/chen20s.html

    2020

  3. [4]

    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 2009 IEEE Conference on Computer Vision and Pattern Recognition , pp.\ 248--255, June 2009. doi:10.1109/CVPR.2009.5206848. URL https://ieeexplore.ieee.org/document/5206848. ISSN: 1063-6919

    arXiv 2009

  4. [5]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan (eds.), Advances in Neural Information Processing Systems, volume 34, pp.\ 8780--8794. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper_files/paper/2021/file/49ad23d1ec9...

    2021

  5. [6]

    PixelAsParam : A Gradient View on Diffusion Sampling with Guidance

    Anh-Dung Dinh, Daochang Liu, and Chang Xu. PixelAsParam : A Gradient View on Diffusion Sampling with Guidance . In Proceedings of the 40th International Conference on Machine Learning , pp.\ 8120--8137. PMLR, July 2023 a . URL https://proceedings.mlr.press/v202/dinh23a.html. ISSN: 2640-3498

    2023

  6. [7]

    Rethinking conditional diffusion sampling with progressive guidance

    Anh-Dung Dinh, Daochang Liu, and Chang Xu. Rethinking conditional diffusion sampling with progressive guidance. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems, volume 36, pp.\ 42285--42297. Curran Associates, Inc., 2023 b . URL https://proceedings.neurips.cc/paper_files/pape...

    2023

  7. [8]

    Representative Guidance : Diffusion Model Sampling with Coherence

    Anh-Dung Dinh, Daochang Liu, and Chang Xu. Representative Guidance : Diffusion Model Sampling with Coherence . In The Thirteenth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=gWgaypDBs8

    2024

  8. [9]

    Endres and J.E

    D.M. Endres and J.E. Schindelin. A new metric for probability distributions. IEEE Transactions on Information Theory, 49 0 (7): 0 1858--1860, July 2003. ISSN 1557-9654. doi:10.1109/TIT.2003.813506. URL https://ieeexplore.ieee.org/document/1207388/

    arXiv 2003

  9. [10]

    Taming Transformers for High-Resolution Image Synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis . In 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pp.\ 12868--12878. IEEE, 2021. ISBN 978-1-6654-4509-2. doi:10.1109/CVPR46437.2021.01268. URL https://ieeexplore.ieee.org/document/9578911/

    arXiv 2021

  10. [11]

    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 . In The Thirteenth International Conference on Learning Representations , 2024. URL https://openreview.net/forum?id=jQP5o1VAVc

    2024

  11. [12]

    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. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems, volume 36, pp.\ 52132--52152. Curran Associates, Inc., 2023. URL https://proceedings.neurips.c...

    2023

  12. [13]

    Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis . In 2025 IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , 2024. doi:10.48550/arXiv.2412.04431. URL http://arxiv.org/abs/2412.04431

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.04431 2025

  13. [14]

    Classifier- Free Diffusion Guidance , 2022

    Jonathan Ho and Tim Salimans. Classifier- Free Diffusion Guidance , 2022. URL http://arxiv.org/abs/2207.12598

    Pith/arXiv arXiv 2022

  14. [15]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models . In Advances in Neural Information Processing Systems , volume 33, pp.\ 6840--6851. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/hash/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.html

    2020

  15. [17]

    Guiding a Diffusion Model with a Bad Version of Itself , December 2024 a

    Tero Karras, Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a Diffusion Model with a Bad Version of Itself , December 2024 a . URL http://arxiv.org/abs/2406.02507. arXiv:2406.02507 [cs]

    Pith/arXiv arXiv 2024

  16. [18]

    Analyzing and Improving the Training Dynamics of Diffusion Models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and Improving the Training Dynamics of Diffusion Models . In 2024 IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pp.\ 24174--24184, Seattle, WA, USA, June 2024 b . IEEE. ISBN 979-8-3503-5300-6. doi:10.1109/CVPR52733.2024.02282. ...

    arXiv 2024

  17. [19]

    YOLOv11 : An Overview of the Key Architectural Enhancements , October 2024

    Rahima Khanam and Muhammad Hussain. YOLOv11 : An Overview of the Key Architectural Enhancements , October 2024. URL http://arxiv.org/abs/2410.17725. arXiv:2410.17725 [cs]

    Pith/arXiv arXiv 2024

  18. [20]

    Pick-a- Pic : An Open Dataset of User Preferences for Text -to- Image Generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a- Pic : An Open Dataset of User Preferences for Text -to- Image Generation . In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems , volume 36, pp.\ 36652--36663. Curran Associates, Inc.,...

    2023

  19. [21]

    Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models

    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 . In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems , volume 37, pp....

    2024

  20. [22]

    Playground v2.5: Three Insights towards Enhancing Aesthetic Quality in Text -to- Image Generation , February 2024 a

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three Insights towards Enhancing Aesthetic Quality in Text -to- Image Generation , February 2024 a . URL http://arxiv.org/abs/2402.17245. arXiv:2402.17245 [cs]

    Pith/arXiv arXiv 2024

  21. [23]

    Adaptive Classifier - Free Guidance via Dynamic Low - Confidence Masking , May 2025

    Pengxiang Li, Shilin Yan, Joey Tsai, Renrui Zhang, Ruichuan An, Ziyu Guo, and Xiaowei Gao. Adaptive Classifier - Free Guidance via Dynamic Low - Confidence Masking , May 2025. URL http://arxiv.org/abs/2505.20199. arXiv:2505.20199 [cs]

    Pith/arXiv arXiv 2025

  22. [24]

    Autoregressive Image Generation without Vector Quantization

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive Image Generation without Vector Quantization . In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems , volume 37, pp.\ 56424--56445. Curran Associates, Inc., 2024 b . URL https://proceedi...

    2024

  23. [25]

    Reward- Guided Speculative Decoding for Efficient LLM Reasoning , June 2025

    Baohao Liao, Yuhui Xu, Hanze Dong, Junnan Li, Christof Monz, Silvio Savarese, Doyen Sahoo, and Caiming Xiong. Reward- Guided Speculative Decoding for Efficient LLM Reasoning , June 2025. URL http://arxiv.org/abs/2501.19324. arXiv:2501.19324 [cs]

    Pith/arXiv arXiv 2025

  24. [26]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO : Common Objects in Context . In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (eds.), Computer Vision – ECCV 2014 , pp.\ 740--755, Cham, 2014. Springer International Publishing. ISBN 978-3-319-10602-...

    doi:10.1007/978-3-319-10602-1_48 2014

  25. [27]

    Coherent Semantic Attention for Image Inpainting

    Hongyu Liu, Bin Jiang, Yi Xiao, and Chao Yang. Coherent Semantic Attention for Image Inpainting . In 2019 IEEE / CVF International Conference on Computer Vision ( ICCV ) , pp.\ 4169--4178, Seoul, Korea (South), October 2019. IEEE. ISBN 978-1-7281-4803-8. doi:10.1109/ICCV.2019.00427. URL https://ieeexplore.ieee.org/document/9009473/

    arXiv 2019

  26. [28]

    Unlocking Guidance for Discrete State - Space Diffusion and Flow Models , March 2025

    Hunter Nisonoff, Junhao Xiong, Stephan Allenspach, and Jennifer Listgarten. Unlocking Guidance for Discrete State - Space Diffusion and Flow Models , March 2025. URL http://arxiv.org/abs/2406.01572. arXiv:2406.01572 [cs]

    Pith/arXiv arXiv 2025

  27. [29]

    Scalable Diffusion Models with Transformers

    William Peebles and Saining Xie. Scalable Diffusion Models with Transformers . In 2023 IEEE / CVF International Conference on Computer Vision ( ICCV ) , pp.\ 4172--4182, 2023. doi:10.1109/ICCV51070.2023.00387. URL https://ieeexplore.ieee.org/document/10377858

    arXiv 2023

  28. [30]

    E. C. Pielou. The measurement of diversity in different types of biological collections. Journal of Theoretical Biology, 13: 0 131--144, December 1966. ISSN 0022-5193. doi:10.1016/0022-5193(66)90013-0. URL https://www.sciencedirect.com/science/article/pii/0022519366900130

    arXiv 1966

  29. [31]

    SDXL : Improving Latent Diffusion Models for High - Resolution Image Synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL : Improving Latent Diffusion Models for High - Resolution Image Synthesis . In The Twelfth International Conference on Learning Representations, October 2023. URL https://openreview.net/forum?id=di52zR8xgf

    2023

  30. [32]

    Zero- Shot Text-to-Image Generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero- Shot Text-to-Image Generation . In Proceedings of the 38th International Conference on Machine Learning , pp.\ 8821--8831. PMLR, 2021. URL https://proceedings.mlr.press/v139/ramesh21a.html

    2021

  31. [33]

    FlowAR : Scale -wise Autoregressive Image Generation Meets Flow Matching , December 2024

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. FlowAR : Scale -wise Autoregressive Image Generation Meets Flow Matching , December 2024. URL http://arxiv.org/abs/2412.15205. arXiv:2412.15205 [cs]

    Pith/arXiv arXiv 2024

  32. [34]

    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 2022 IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pp.\ 10674--10685, 2022. doi:10.1109/CVPR52688.2022.01042. URL https://ieeexplore.ieee.org/document/9878449

    arXiv 2022

  33. [35]

    Michael S. Ryoo, A. J. Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. TokenLearner : What Can 8 Learned Tokens Do for Images and Videos ?, April 2022. URL http://arxiv.org/abs/2106.11297. arXiv:2106.11297 [cs]

    Pith/arXiv arXiv 2022

  34. [36]

    de Almeida, Alexander Rush, Thomas Pierrot, and Volodymyr Kuleshov

    Yair Schiff, Subham Sekhar Sahoo, Hao Phung, Guanghan Wang, Sam Boshar, Hugo Dalla-torre, Bernardo P. de Almeida, Alexander Rush, Thomas Pierrot, and Volodymyr Kuleshov. Simple Guidance Mechanisms for Discrete Diffusion Models , May 2025. URL http://arxiv.org/abs/2412.10193. arXiv:2412.10193 [cs]

    Pith/arXiv arXiv 2025

  35. [37]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models . In The Eighth International Conference on Learning Representations , 2020. URL https://openreview.net/forum?id=St1giarCHLP

    2020

  36. [38]

    Autoregressive Model Beats Diffusion : Llama for Scalable Image Generation , June 2024

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive Model Beats Diffusion : Llama for Scalable Image Generation , June 2024. URL http://arxiv.org/abs/2406.06525. arXiv:2406.06525 [cs]

    Pith/arXiv arXiv 2024

  37. [39]

    HART : Efficient Visual Generation with Hybrid Autoregressive Transformer , 2024

    Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. HART : Efficient Visual Generation with Hybrid Autoregressive Transformer , 2024. URL https://openreview.net/forum?id=q5sOv4xQe4

    2024

  38. [40]

    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 . In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems , volume 37, pp.\ 84839--84865. Curran Associates, Inc., 2024. ...

    2024

  39. [41]

    Switti: Designing Scale-Wise Transformers for Text-to-Image Synthesis , 2025

    Anton Voronov, Denis Kuznedelev, Mikhail Khoroshikh, Valentin Khrulkov, and Dmitry Baranchuk. Switti: Designing Scale-Wise Transformers for Text-to-Image Synthesis , 2025. URL http://arxiv.org/abs/2412.01819

    Pith/arXiv arXiv 2025

  40. [42]

    ImageReward : Learning and Evaluating Human Preferences for Text -to- Image Generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward : Learning and Evaluating Human Preferences for Text -to- Image Generation . In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neural Information Processing Systems , volume 36, pp.\ 15903--15935. Curran Asso...

    2023

  41. [43]

    Image Captioning With Semantic Attention

    Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image Captioning With Semantic Attention . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 4651--4659, 2016. URL https://openaccess.thecvf.com/content_cvpr_2016/html/You_Image_Captioning_With_CVPR_2016_paper.html

    2016

  42. [44]

    How Much To Guide : Revisiting Adaptive Guidance in Classifier - Free Guidance Text -to- Vision Diffusion Models , June 2025

    Huixuan Zhang, Junzhe Zhang, and Xiaojun Wan. How Much To Guide : Revisiting Adaptive Guidance in Classifier - Free Guidance Text -to- Vision Diffusion Models , June 2025. URL http://arxiv.org/abs/2506.08351. arXiv:2506.08351 [cs]

    Pith/arXiv arXiv 2025

  43. [45]

    VAR-CLIP : Text-to-Image Generator with Visual Auto-Regressive Modeling , 2024

    Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. VAR-CLIP : Text-to-Image Generator with Visual Auto-Regressive Modeling , 2024. URL http://arxiv.org/abs/2408.01181

    Pith/arXiv arXiv 2024

  44. [46]

    Entropy- Driven Sampling and Training Scheme for Conditional Diffusion Generation

    Guangcong Zheng, Shengming Li, Hui Wang, Taiping Yao, Yang Chen, Shouhong Ding, and Xi Li. Entropy- Driven Sampling and Training Scheme for Conditional Diffusion Generation . In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner (eds.), Computer Vision – ECCV 2022 , volume 13682, pp.\ 754--769. Springer Nature Switzer...

    work page doi:10.1007/978-3-031-20047-2_43 2022

  45. [47]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  46. [48]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  47. [49]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...