Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Ordering image tokens by the breadth-first traversal of a random spanning tree lets a standard causal autoregressive transformer keep raster-scan-level sampling quality while enabling connected-mask image inpainting at inference time.

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 ↗

Breadth-first traversal of random spanning trees as token order keeps autoregressive image quality and enables connected-mask inpainting without changing the transformer architecture.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection STAR is a genuine new trick for random-order AR image generation—BFS traversal of uniform spanning trees—but its editing claim is narrower than advertised because the postfix completion guarantee silently requires the mask to be connected, which the paper's own evaluation always ensures. the 3 major comments →

arxiv 2511.17089 v2 pith:JFLXYITO submitted 2025-11-21 cs.CV cs.AI

Spanning Tree Autoregressive Visual Generation

Sangkyu Lee , Changho Lee , Janghoon Han , Hosung Song , Tackgeun You , Hwasup Lim , Stanley Jungkyu Choi , Honglak Lee
show 1 more author
Youngjae Yu
This is my paper
classification cs.CV cs.AI
keywords autoregressive image generationsequence order randomizationuniform spanning treebreadth-first traversalimage inpaintingnext-position embeddingscausal transformerclass-conditional generation
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.

The reading

The paper attempts to show that a causal autoregressive transformer for image generation can retain the sampling quality of the usual raster-scan order while gaining the freedom to choose the generation order at inference. The proposed mechanism, STAR, replaces the fixed raster-scan order and the fully random permutation order with the breadth-first traversal order of a uniform random spanning tree on the grid of image patch positions. Because these trees are sampled randomly but respect local adjacency and grow from a corner, the model is exposed to bidirectional context without the sample-space explosion of random permutations. The practical payoff is image editing: when the observed part of an image is a connected patch region that does not touch all four corners, STAR's inference algorithm makes the observed region the prefix of the sequence and the masked region the postfix. Empirically, STAR matches or nearly matches raster-scan FID on ImageNet 256x256 and substantially outperforms a randomized-order baseline on connected-mask inpainting.

Core claim

STAR shows that the sequence order used to train an autoregressive visual model can be modeled as a breadth-first traversal of a uniform spanning tree of the token-position lattice, with an extra positional embedding telling the model which token comes next. This single change gives the model the same ability to see context from both sides as random-permutation training, while preserving the local continuity and corner-to-center progress that make raster-scan orders easy to learn. On class-conditional ImageNet generation, STAR-XXL reaches FID 1.55, close to the 1.48 of RAR-XXL and far better than random-order baselines; on inpainting with connected masks, STAR outperforms RAR across mask rat

What carries the argument

The central object is the uniform spanning tree on the h×w grid lattice whose vertices are image patch positions. The paper samples such a tree (via Wilson's loop-erased random walk), roots it at a random corner, and uses its BFS order as the token sequence. For inpainting, Algorithm 1 uses rejection sampling to find a tree of the unmasked region whose maximum-depth vertices lie on the mask boundary, so the masked tokens can be appended as a postfix; the next-position embedding tells the decoder which grid cell to predict.

Load-bearing premise

The postfix-completion guarantee assumes the observed (unmasked) part of the image is a single connected block of lattice cells and that the mask does not cover every corner; the paper evaluates editing only on masks generated to satisfy this precondition.

What would settle it

Take a 16×16 lattice and a connected mask whose complement is a long, narrow diagonal strip that touches two opposite corners — both properties allowed by the paper's preconditions. Run Algorithm 1 with the prescribed BFS rejection sampling; if the acceptance rate is not near 1 within a few hundred trials, the postfix-completion guarantee is not structural.

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

If this is right

  • STAR-XXL reaches FID 1.55 on ImageNet 256×256 class-conditional generation, comparable to RAR-XXL's 1.48 and well ahead of random-permutation models like RandAR, without adding an order-annealing schedule or changing the decoder-only architecture.
  • For connected-mask inpainting with mask ratios 0.1 to 0.9, STAR outperforms the random-permutation baseline RAR at every model scale, with the largest gains where roughly half the image is observed.
  • BFS traversal of the spanning tree makes the inference-time rejection sampling in Algorithm 1 succeed in at most a handful of trials (average under 6, failure rate 0.0 in the tested range), whereas DFS fails on a substantial fraction of masks.
  • The conditional entropy of STAR-trained models stays roughly uniform along the sequence, mirroring raster-scan training; the paper connects this uniformity to the superior sampling performance compared with random orders.

Where Pith is reading between the lines

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

  • Because the postfix-completion guarantee requires the observed region to be connected and not to touch all four corners, real-world editing masks that are scattered or border-spanning are outside the demonstrated capability; a natural extension is to split the observed region into connected components and chain them, which the paper does not study.
  • The paper's entropy analysis suggests the benefit comes from keeping prediction difficulty uniform; one testable consequence is that any family of trees whose BFS orders have similar locality and corner-to-center structure — for example, minimum spanning trees under a distance-dependent edge weight — would yield similar gains.
  • Since STAR only changes the order and the positional conditioning, the recipe should transfer to other causal AR architectures and continuous tokenizers without retraining the image tokenizer; testing this would show whether the mechanism is tokenizer-independent.
  • The mask-generation algorithm in the supplement always preserves connectivity of both the mask and its complement; evaluating on an unbiased sample of all connected masks would clarify whether acceptance and FID gains hold beyond the favorable mask distribution used in the paper.
Share X Bluesky LinkedIn Reddit HN

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 introduces Spanning Tree Autoregressive (STAR) modeling for decoder-only AR visual generation. Instead of a fixed raster-scan order or a uniformly random permutation, STAR samples a uniform spanning tree on the h×w patch lattice, chooses a root uniformly among the four corners, and uses the BFS traversal order as the token sequence, providing next-position embeddings as conditions. The training objective in Eq. (4) marginalizes over roots and trees. For inpainting, Algorithm 1 uses rejection sampling to construct a tree whose BFS order places a connected unmasked region as a prefix and the masked region as a postfix, enabling postfix completion. On ImageNet 256, STAR-B/L/XL/XXL achieve FIDs 2.24/1.98/1.65/1.55 (Table 1), outperforming RandAR and roughly matching RAR at larger scales with higher IS. Inpainting results (Fig. 6, Table 2) show STAR outperforming RAR under the connected-mask evaluation protocol. Ablations compare random permutation, BFS/DFS traversal, and root-selection strategies.

Significance. The core idea is original and simple: restrict the order-randomization space from N! permutations to the exponentially large but structured space of spanning-tree BFS orders. This is well motivated by the entropy measurements (Figs. 3–4) and by the locality/center-bias priors. The paper provides replication-friendly hyperparameters (Supp. Table A), a public code release, and a clean ablation structure (Tables 2–3). If the claims hold, STAR offers a practical way to give causal AR image models inference-time order flexibility without architecture changes. The main caveat is that the demonstrated editing capability is limited to a specific mask class, as discussed below.

major comments (3)
  1. [§3.3, Algorithm 1 (line 16); Supp. Algorithm A] Algorithm 1's Ensure line lists M⊂G, corner(G)⊄M, and G\M connected, but it does not state that M itself must be connected. Yet line 16 samples T(M,r_M), a uniform spanning tree of M, which is undefined when M is disconnected. Condition (3) is not categorically unsatisfiable for disconnected M (two masked components can be attached at equal depth), but the algorithm as written cannot handle multiple mask components. Supp. Algorithm A explicitly generates masks with both M and G\M connected, so the inpainting numbers in Fig. 6 and Table 2 cover only a favorable mask class. The abstract's 'image editing' claim therefore overstates the demonstrated capability. Please state the connected-mask assumption explicitly, extend the algorithm to disconnected masks (e.g., per-component trees or a spanning forest), evaluate on general masks, or temper the claims.
  2. [§4.2, Figure 6] The RAR inpainting protocol is not described precisely. The sentence 'through teacher forcing without sampling from the model for RAR' is ambiguous: if RAR was evaluated by likelihood/teacher-forcing rather than by stochastic sampling with a comparable CFG scale and temperature, the FID/IS gap in Fig. 6 may reflect decoding choices rather than the order model. Please report the exact RAR inference procedure (order construction, CFG scale/power or none, temperature, number of samples, tokenizer) and, ideally, include a STAR variant decoded greedily without CFG to isolate the effect of sequence-order flexibility.
  3. [Abstract; §4.2, Table 1] The claim that STAR 'maintains sampling performance' relative to randomized-order approaches needs qualification. In Table 1, RAR-B/L/XL/XXL has lower FID than STAR at every scale (e.g., 1.48 vs 1.55 at XXL), although STAR shows better inpainting capability. The current wording in the abstract and contribution list implies a strict Pareto improvement over random-order methods. Suggest stating the comparison precisely: STAR is competitive with raster-scan and RAR on generation while adding postfix-completion behavior, and it outperforms RandAR and simple random permutation.
minor comments (6)
  1. [§3.2, Figures 3–4] The axis labels in Figures 3 and 4 appear garbled ('(N) [ (x i)]' and '(N) [ (x i x < i)]'). Please render them properly.
  2. [§3.3, Eq. (3)] There is a typo: 'vm' should be 'v_M'. Also, the notation d_1 is used but not defined at first use; please define it as Manhattan distance.
  3. [§4.1] The 'ten-crop transformation [34,44]' is ambiguous for 256×256 training. Specify whether this is a random-crop augmentation during training or a ten-crop evaluation, and clarify how it is applied to the patch lattice.
  4. [§4.2] For the inpainting FID/IS, the reference distribution is not specified. State whether FID is computed between inpainted validation images and original validation images, or between inpainted and generated unconditonal samples, and report the number of images used.
  5. [§4.3, Table 3] The column header '# Average Trial for Accept (Failure Ratio)' is ambiguous. Clarify that 'Failure Ratio' is the fraction of masks for which no accepting trial was found within the 100-trial bound.
  6. [§5] The conclusion would benefit from a brief limitations paragraph acknowledging the connected-mask restriction for inpainting and the modest FID tradeoff against RAR shown in Table 1.

Circularity Check

0 steps flagged

No significant circularity: STAR's training objective and postfix-completion construction are not fitted to benchmark targets; only a mask-connectedness scope limitation is noted.

full rationale

The paper's derivation chain is self-contained. The training objective (Eq. 4) maximizes likelihood over roots and BFS-ordered uniform spanning trees; it is a well-defined order-randomization distribution and is not fitted to the measured FID/IS/inpainting targets. The postfix-completion guarantee in Algorithm 1 is an algorithmic construction derived from BFS depth properties (Eq. 3), and Table 3 independently ablates BFS vs. DFS acceptance. Generation quality is evaluated against external baselines (LlamaGen, RAR, RandAR, VAR, MAR, diffusion models) on ImageNet-1k, and the only tuned hyperparameters are conventional inference settings (CFG scale/power, temperature). No load-bearing result is imported from a self-citation chain: Wilson's algorithm [41], lattice spanning-tree asymptotics [31], and next-position-embedding ideas [27,44] are external or standard and do not assume STAR's target claims. The one caveat worth flagging is a scope limitation, not circularity: Algorithm 1's Ensure line states 'M⊂G, corner(G)⊄M, G\M is connected' but omits connectedness of M itself, even though line 16 samples a spanning tree T(M,r_M), which is undefined for disconnected M. The inpainting evaluation explicitly restricts to masks where 'both the masked and non-masked regions remain connected' (Supplementary Algorithm A), so the demonstrated postfix-completion capability covers only a favorable mask class. This narrows the editing claim but does not make any prediction equivalent to its input by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

No new physical or formal entities are introduced. The central design rests on empirical image-prior assumptions and on a favorable mask geometry for the editing guarantee; the only numerical free parameters are inference-time hyperparameters.

free parameters (2)
  • Classifier-free guidance scale and power (per model size) = B: 16.0/2.75; L: 14.5/2.55; XL: 6.5/1.5; XXL: 5.4/1.25
    Tuned on ImageNet validation to optimize FID/IS; values differ by model size (Supplementary Table A). These affect all reported generation numbers.
  • Sampling temperature = 0.98
    Fixed inference hyperparameter used for all STAR models; affects FID and diversity.
axioms (6)
  • domain assumption Predicting adjacent/connected patches before distant patches encodes useful image prior; locality and center bias survive in VQ-token space.
    Motivates BFS spanning-tree orders. Supported only by entropy measurements on one random-order model (Section 3.2, Figures 3-4), not by a first-principles argument.
  • domain assumption Sampling sequence orders from the uniform spanning tree distribution exposes the model to sufficient bidirectional context compared with random permutations.
    Central hypothesis tested by ablations (Table 2), but not proven; STAR still underperforms RAR on FID, so 'sufficiency' is only approximate.
  • standard math Wilson's algorithm samples exactly from the uniform spanning tree distribution.
    Used for training-time and inference-time tree sampling (Section 4.1); standard result, accepted without proof.
  • standard math For postfix completion, it is enough that some maximum-depth vertex of the unmasked tree is adjacent to the masked region (Eq. 3).
    Graph-theoretic condition used by Algorithm 1; justified by the BFS traversal argument and empirically by Table 3, but not formally proved.
  • ad hoc to paper Masked region is connected and does not contain all corners; unmasked region remains connected.
    Algorithm 1's 'Require' line restricts image editing to a favorable mask class; the paper's random masking strategy (Supplementary Algorithm A) generates exactly such masks.
  • standard math Next-token positional embeddings suffice to condition an AR transformer on arbitrary sequence orders.
    Borrowed from sigma-GPT and RAR; used for all randomized-order training (Section 4.1).

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Spanning Tree Autoregressive Visual Generation." pith.science (2026). https://pith.science/paper/JFLXYITO

@misc{pith2026251117089,
  author       = {Pith},
  title        = {Pith review of: Spanning Tree Autoregressive Visual Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JFLXYITO}},
  note         = {Machine review of arXiv:2511.17089}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present Spanning Tree Autoregressive (STAR) modeling, which can incorporate prior knowledge of images, such as center bias and locality, to maintain sampling performance while also providing sufficiently flexible sequence orders to accommodate image editing at inference time. Approaches that expose conventional autoregressive (AR) models in visual generation to arbitrary sequence orders via random permutation suffer from degraded sampling performance or compromise the flexibility in sequence order choice at inference time. Instead, STAR utilizes traversal orders of uniform spanning trees in a lattice defined by the positions of image patches. Traversal orders are obtained via breadth-first search, allowing us to efficiently construct a spanning tree via rejection sampling whose traversal order ensures that the connected partial observation of the image appears as a prefix for native image inpainting support. Through the tailored yet structured sequence order randomization strategy, STAR preserves the capability of postfix completion while maintaining sampling performance, without any significant changes to the model architecture widely adopted in language AR modeling.

Figures

Figures reproduced from arXiv: 2511.17089 by Changho Lee, Honglak Lee, Hosung Song, Hwasup Lim, Janghoon Han, Sangkyu Lee, Stanley Jungkyu Choi, Tackgeun You, Youngjae Yu.

Figure 1
Figure 1. Figure 1: Comparison of sequence orders in AR visual generation. (a) Conventional AR models follow a fixed raster-scan order, which [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed STAR modeling. (a) We perform training and inference according to the sequence order obtained [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prediction entropy of tokens according to token position [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of sampled results from STAR-XXL in the class-conditional image generation on ImageNet-1k at [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Quantitative comparison of RAR [44] and STAR on the inpainting task with varying masking ratio from 0.1 to 0.9. Original RAR-XXL STAR-XXL [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on the inpainting task, demonstrating insufficient results of RAR-XXL [ [PITH_FULL_IMAGE:figures/full_fig_p007_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

44 extracted references · 11 linked inside Pith

  1. [1]

    All are worth words: A vit backbone for diffusion models

    Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 22669–22679, 2023. 1, 2, 6

    2023

  2. [2]

    Halton scheduler for masked gen- erative image transformer.arXiv preprint arXiv:2503.17076,

    Victor Besnier, Mickael Chen, David Hurych, Eduardo Valle, and Matthieu Cord. Halton scheduler for masked gen- erative image transformer.arXiv preprint arXiv:2503.17076,

    Pith/arXiv arXiv

  3. [3]

    Reconciling saliency and object center-bias hypotheses in explaining free-viewing fixations

    Ali Borji and James Tanner. Reconciling saliency and object center-bias hypotheses in explaining free-viewing fixations. IEEE transactions on neural networks and learning systems, 27(6):1214–1226, 2015. 2, 4

    2015

  4. [4]

    Lan- guage models are few-shot learners.Advances in neural in- formation processing systems, 33:1877–1901, 2020

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Lan- guage models are few-shot learners.Advances in neural in- formation processing systems, 33:1877–1901, 2020. 2

    1901

  5. [5]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 11315–11325, 2022. 1, 2, 5

    2022

  6. [6]

    Generative pre- training from pixels

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Hee- woo Jun, David Luan, and Ilya Sutskever. Generative pre- training from pixels. InInternational conference on machine learning, pages 1691–1703. PMLR, 2020. 1, 2, 3

    2020

  7. [7]

    Scaling vision transformers to 22 billion pa- rameters

    Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul- mohsin, et al. Scaling vision transformers to 22 billion pa- rameters. InInternational conference on machine learning, pages 7480–7512. PMLR, 2023. 5

    2023

  8. [8]

    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. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 4, 5

    2009

  9. [9]

    Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 6

    2021

  10. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 5

    Pith/arXiv arXiv 2010

  11. [11]

    Scalable pre- training of large autoregressive image models

    Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Vaishaal Shankar, Alexander To- shev, Joshua M Susskind, and Armand Joulin. Scalable pre- training of large autoregressive image models. InProceed- ings of the 41st International Conference on Machine Learn- ing, pages 12371–12384, 2024. 1, 3

    2024

  12. [12]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021. 2, 3, 5

    2021

  13. [13]

    Masked diffusion transformer is a strong image synthesizer

    Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Masked diffusion transformer is a strong image synthesizer. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 23164–23173,

  14. [14]

    Scaling laws for autoregressive generative modeling.arXiv preprint arXiv:2010.14701, 2020

    Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling.arXiv preprint arXiv:2010.14701, 2020. 1, 2

    Pith/arXiv arXiv 2010

  15. [15]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

    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 equilib- rium.Advances in neural information processing systems, 30, 2017. 6

    2017

  16. [16]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 6

    Pith/arXiv arXiv 2022

  17. [17]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2

    2020

  18. [18]

    Statistics of natural images and models

    Jinggang Huang and David Mumford. Statistics of natural images and models. InProceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recog- nition (Cat. No PR00149), pages 541–547. IEEE, 1999. 2, 4

    1999

  19. [19]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,

    Pith/arXiv arXiv 2001

  20. [20]

    Computational tradeoffs in image synthesis: Diffusion, masked-token, and next-token prediction.arXiv preprint arXiv:2405.13218, 2024

    Maciej Kilian, Varun Jampani, and Luke Zettlemoyer. Computational tradeoffs in image synthesis: Diffusion, masked-token, and next-token prediction.arXiv preprint arXiv:2405.13218, 2024. 1, 2

    Pith/arXiv arXiv 2024

  21. [21]

    Improved precision and recall met- ric for assessing generative models.Advances in neural in- formation processing systems, 32, 2019

    Tuomas Kynk ¨a¨anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall met- ric for assessing generative models.Advances in neural in- formation processing systems, 32, 2019. 6

    2019

  22. [22]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. InProceedings of the IEEE/CVF con- 12 ference on computer vision and pattern recognition, pages 11523–11532, 2022. 2

    2022

  23. [23]

    Au- toregressive image generation with randomized parallel de- coding.arXiv preprint arXiv:2503.10568, 2025

    Haopeng Li, Jinyue Yang, Guoqi Li, and Huan Wang. Au- toregressive image generation with randomized parallel de- coding.arXiv preprint arXiv:2503.10568, 2025. 1, 2, 3, 7

    arXiv 2025

  24. [24]

    Autoregressive image generation without vec- tor quantization.Advances in Neural Information Processing Systems, 37:56424–56445, 2024

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vec- tor quantization.Advances in Neural Information Processing Systems, 37:56424–56445, 2024. 6

    2024

  25. [25]

    Elucidating the de- sign space of language models for image generation.arXiv preprint arXiv:2410.16257, 2024

    Xuantong Liu, Shaozhe Hao, Xianbiao Qi, Tianyang Hu, Jun Wang, Rong Xiao, and Yuan Yao. Elucidating the de- sign space of language models for image generation.arXiv preprint arXiv:2410.16257, 2024. 1, 2

    Pith/arXiv arXiv 2024

  26. [26]

    Randar: Decoder-only autoregressive visual generation in random orders

    Ziqi Pang, Tianyuan Zhang, Fujun Luan, Yunze Man, Hao Tan, Kai Zhang, William T Freeman, and Yu-Xiong Wang. Randar: Decoder-only autoregressive visual generation in random orders. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 45–55, 2025. 1, 2, 3, 6, 7

    2025

  27. [27]

    InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 143–159

    Arnaud Pannatier, Evann Courdier, and Franc ¸ois Fleuret.σ- gpts: A new approach to autoregressive models. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 143–159. Springer, 2024. 1, 2, 5, 6

    2024

  28. [28]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205,

  29. [29]

    Language models are unsu- pervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsu- pervised multitask learners.OpenAI blog, 1(8):9, 2019. 1, 2

    2019

  30. [30]

    Improved techniques for training gans.Advances in neural information processing systems, 29, 2016

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans.Advances in neural information processing systems, 29, 2016. 6

    2016

  31. [31]

    Spanningtrees on graphs and lattices in d dimensions.Journal of Physics A: Mathe- matical and General, 33(21):3881, 2000

    Robert Shrock and Fa Yueh Wu. Spanningtrees on graphs and lattices in d dimensions.Journal of Physics A: Mathe- matical and General, 33(21):3881, 2000. 4

    2000

  32. [32]

    Autoregressive model beats diffusion: Llama for scalable image generation.arXiv preprint arXiv:2406.06525, 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.arXiv preprint arXiv:2406.06525, 2024. 2, 6, 7

    Pith/arXiv arXiv 2024

  33. [33]

    Mitigating the bias of centered objects in common datasets

    Gergely Szab ´o and Andr ´as Horv´ath. Mitigating the bias of centered objects in common datasets. In2022 26th Inter- national Conference on Pattern Recognition (ICPR), pages 4786–4792. IEEE, 2022. 2, 4

    2022

  34. [34]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015. 6

    2015

  35. [35]

    Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818, 2024

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818, 2024. 1, 2

    Pith/arXiv arXiv 2024

  36. [36]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural in- formation processing systems, 37:84839–84865, 2024

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Li- wei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural in- formation processing systems, 37:84839–84865, 2024. 2, 6, 7

    2024

  37. [37]

    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, 2

    Pith/arXiv arXiv 2023

  38. [38]

    Pixel recurrent neural networks

    A ¨aron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. InInterna- tional conference on machine learning, pages 1747–1756. PMLR, 2016. 1, 2

    2016

  39. [39]

    Neural discrete representation learning.Advances in neural information pro- cessing systems, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Advances in neural information pro- cessing systems, 30, 2017. 1, 2

    2017

  40. [40]

    Analysis of classifier-free guidance weight schedulers.Transactions on Machine Learning Research Journal, 2024

    Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fern´andez Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers.Transactions on Machine Learning Research Journal, 2024. 6

    2024

  41. [41]

    Generating random spanning trees more quickly than the cover time

    David Bruce Wilson. Generating random spanning trees more quickly than the cover time. InProceedings of the twenty-eighth annual ACM symposium on Theory of com- puting, pages 296–303, 1996. 4, 6

    1996

  42. [42]

    Xlnet: Generalized autoregressive pretraining for language understanding.Ad- vances in neural information processing systems, 32, 2019

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding.Ad- vances in neural information processing systems, 32, 2019. 1, 2

    2019

  43. [43]

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

    Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan.arXiv preprint arXiv:2110.04627, 2021. 2

    Pith/arXiv arXiv 2021

  44. [44]

    Randomized autoregressive visual generation

    Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang- Chieh Chen. Randomized autoregressive visual generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18431–18441, 2025. 1, 2, 3, 5, 6, 7, 8 13

    2025

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.