Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SpectralAR: Spectral Autoregressive Visual Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Ordering image tokens from low to high DCT frequency gives autoregressive generation a true causal sequence, reaching 3.02 gFID with only 64 tokens and 310M parameters.

desk verdict Solid empirical paper with a plausible but under-supported causal story; worth serious review. read the letter →

arxiv 2506.10962 v1 pith:KOK7SOFP submitted 2025-06-12 cs.CV cs.AIcs.LG

classification cs.CVcs.AIcs.LG
keywords spectralautoregressivegenerationdiscretecosinetransformnestedtokenizationtokenefficiencycausalvisualsequenceclass-conditionalimageNet-1K
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the reason autoregressive image generation lags behind diffusion is that spatial patch order is not a genuine causal sequence: adjacent patches are mutually independent, violating the next-token premise. It proposes representing an image as a 1D sequence of spectral tokens ordered from low to high DCT frequency components, where each new token adds a higher-frequency band of detail, and generating images by predicting those tokens in order. With this order and a nested tokenization that reuses earlier tokens, the model reaches 3.02 gFID on ImageNet-1K with only 64 tokens and 310M parameters, outperforming the autoregressive TiTok baseline (6.30) and VAR-d16 (3.30) at similar scale. If correct, this means the ordering of visual information, not a larger model, is what makes causal visual generation work.

What carries the argument

Nested Spectral Tokenization (NST) is the central object. It maps an image to DCT coefficients, defines a sequence of frequency thresholds $\{\omega_i\}$, and constructs sub-images $I'_i = \text{DCT}^{-1}(D \circ \mathbb{1}_{\omega_i})$, where $\mathbb{1}_{\omega_i}$ keeps only the top-left $\omega_i \times \omega_i$ coefficient block. The sequence of $N$ tokens is supervised so that token $s_i$ (with all previous tokens) reconstructs $I'_i$, making the representation nested: tokens are reused across levels rather than split into independent groups. This containment property provides the causal structure and allows a 1D causal mask in both tokenizer encoder and decoder, and the non-uniform spacing of $\omega_i$ (more tokens for low, fewer for high frequencies) is what makes 64 tokens sufficient.

What would settle it

Train the identical generator and tokenizer but present the spectral tokens to the autoregressive model in a randomly shuffled order (or in reverse frequency order) while keeping all other training details fixed. If the spectral ordering itself is causally beneficial, the shuffled model should produce substantially worse gFID than the reported 3.02; if FID stays close to 3.02, the ordering is not the driver of the gain.

Watch

Extended reading notes

Core claim

The central claim is that images admit a causal 1D sequence in the spectral domain: transform the image by DCT, then define nested sub-images by keeping only coefficients in an ever-growing low-frequency square $\omega_i \times \omega_i$. Each spectral token $s_i$ is trained so that the decoder, given tokens $s_1,\dots,s_i$, reconstructs sub-image $I'_i$; because $I'_i$ contains all frequencies of $I'_{i-1}$, the tokens nest, so later tokens only add higher-frequency detail. The paper argues that this coarse-to-fine containment is the natural causal order for images, analogous to human perception and painting, and supports it with a linear-correlation proxy on CIFAR-100 where spectral sequences reach 0.962 average R², versus 0.454 for raster-scan and 0.928 for scale-wise sequences. With a non-uniform token-frequency mapping that spends more tokens on low frequencies, plus a causal mask in the tokenizer and a standard GPT-2-like generator, SpectralAR achieves 3.02 gFID with 64 tokens at 310M parameters on ImageNet-1K class-conditional generation.

Load-bearing premise

The paper assumes that the fixed low-to-high DCT frequency order is the order in which image content is genuinely predictable, so that knowing the coarse low-frequency tokens makes higher-frequency tokens conditionally easy to generate; the evidence for this is a linear correlation statistic on CIFAR-100, not a measure of actual generative difficulty.

Editorial extensions

If this is right

  • If the claim holds, visual tokens can be made 1D-causal without sacrificing generation quality, enabling autoregressive visual generation to plug directly into language-model-style next-token frameworks.
  • The 64-token budget means the visual context of multimodal models can be sharply reduced, potentially easing long-context constraints.
  • Truncating the last few tokens degrades quality only slowly (5 tokens: 3.34, 10 tokens: 6.65), giving a controllable quality/efficiency knob.
  • The same tokenizer and generator can be repurposed for super-resolution by refining frequency bands of upsampled sub-images (SpectralAR-Stride: 2.93 FID versus 3.09 for simple upsampling).
  • Larger generators keep improving, with SpectralAR-d24 reaching 2.13 gFID at 1.0B parameters, matching the scaling trend of stronger baselines.

Reading between the lines

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

  • The coarse-to-fine containment property is not unique to DCT: any invertible multiresolution decomposition (wavelet banks, Laplacian pyramids) could in principle yield equally causal token sequences, and comparing those would tell whether the spectral basis itself or the nesting property is what matters.
  • The linear-correlation proxy understates what a generative model needs; a stronger test would be to measure the conditional log-likelihood of high-frequency tokens given low-frequency ones and compare it against the unconditional case.
  • Because the method stays in the standard next-token paradigm, its token representation could transfer to video by applying DCT over the temporal dimension, though the paper does not test this.
  • The rFID of the spectral tokenizer (4.03) is worse than TiTok's (1.70), yet generation improves; this suggests reconstruction quality and sequence causality are partly decoupled, which, if true, would revise how tokenizers are evaluated for autoregressive use.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes SpectralAR, a framework for autoregressive visual generation that represents an image as a 1D sequence of spectral tokens ordered from low to high DCT frequencies. The tokenizer, Nested Spectral Tokenization, supervises nested prefixes of the token sequence against cumulative low-pass reconstructions, uses a non-uniform token-frequency mapping, and applies a causal mask to the spectral tokens in the encoder and decoder. A GPT-like transformer is then trained to predict these spectral tokens in order. On ImageNet-1K 256×256 class-conditional generation, the authors report 3.02 gFID with 64 tokens and 310M parameters, outperforming their TiTok AR baseline (6.30 gFID) and V AR-d16 (3.30) at comparable scale. Ablations in Table 4 attribute the gain to spectral supervision, causal masking, and the non-uniform mapping.

Significance. If the claims are correct, SpectralAR offers a concrete way to make visual sequences genuinely causal while improving token efficiency, addressing a known weakness of spatial-patch autoregressive models. The central idea is clearly formulated, and the ablation study decomposes the contribution of each component. However, the evidence for the spectral-causality premise is indirect: the proof-of-concept correlation metric is confounded by nested support, the comparison omits the closest spectral autoregressive baselines, and the quantitative results are single-run point estimates. With additional controls, the paper could substantiate its main explanatory claim.

major comments (4)
  1. [Section 4.2, Eqs. (5) and (10)] The R²_avg metric is computed between cumulative sub-images I'_i and previous cumulative sub-images, and by construction I'_i contains I'_{i-1} as a low-pass version (Eq. 5). The high values in Table 1 therefore largely reflect shared low-frequency content, not the predictability of the newly added high-frequency band that the next token is supposed to provide. Please measure the predictability of the incremental band D ⊙ (1_{ω_i} − 1_{ω_{i-1}}) (or an equivalent residual) and perform the measurement on the quantized spectral token sequence S rather than raw pixel sub-images.
  2. [Section 3.2, 'Spectral causal mask'] The mask is described as applying to the spectral tokens S, but in Eq. (3) the encoder input is [P; S0], where P are patch tokens from the full image. If P are unmasked, each spectral token can still attend to full-image information from all patches, so the tokenizer sequence is not causal in the claimed sense. Please specify whether P are also causally masked, and if not, provide an ablation (e.g., a variant with masked P) to show that the generation gain is not due to train/test mismatch.
  3. [Section 4.3, Table 2 and Section 2] SIT [11] and CART [43] are the closest spectral autoregressive methods and are only discussed qualitatively. Without benchmarking them, the paper cannot support the claim that the proposed low-to-high DCT ordering is better than existing spectral autoregressive formulations. Moreover, the comparison to TiTok-B64-d16 uses a tokenizer with very different reconstruction quality (rFID 1.70 vs 4.03), which confounds tokenizer training details with sequence causality. Please add comparisons to SIT/CART and a control experiment with reversed or shuffled spectral order.
  4. [Tables 2 and 4] All FID values are point estimates from single training runs, without standard deviations or seeds. The margins between some ablation rows (3.49 vs 3.02) and versus V AR-d16 (3.02 vs 3.30) are small relative to typical run-to-run FID variability on ImageNet-1K; please report results over at least two or three seeds.
minor comments (5)
  1. [Section 3.2, Eq. (9)] 'ωs' is likely a typo for 'ω_i'; please also state the definition as a complete sequence for N=64 tokens.
  2. [Table 3 title] 'SepctralAR' should be 'SpectralAR'.
  3. [Section 4.2, Eq. (10)] Please define the index sets over which AVG and max are taken; as written, the ranges over elements of t_i and the previous tokens are ambiguous.
  4. [Figure 4] The caption and legend are unclear about what is plotted (the 'highest correlation' with which token?) and how the color scale maps to the three methods.
  5. [Section 3.3] 'each spectral tokens s_i is expected to enhance' should be 'each spectral token s_i is expected to enhance'.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; spectral order is evaluated by independent FID benchmarks and ablations.

full rationale

SpectralAR's central claims are empirical: with a fixed, hand-designed spectral schedule (Eq. 9) and nested tokenization, the autoregressive generator reaches 3.02 gFID on ImageNet-1K. Nothing in this chain is fitted to the reported FID: the non-uniform token-frequency mapping (Eqs. 7-9) is derived from the JPEG/compression principle of suppressing high-frequency detail, and the ablation in Table 4 separately validates spectral supervision, causal mask, and non-uniform mapping. The generator uses standard cross-entropy next-token prediction; no parameter of the generator is set by matching the FID. Self-citations [21, 73] appear only as examples of autoregressive world models in related work and are not used to justify any architectural choice or to exclude alternatives. There is no imported uniqueness theorem. The only potentially self-referential evidence is the Section 4.2 proof-of-concept: R^2_avg is computed between nested sub-images I'_i = DCT^{-1}(D ⊙ 1_{ω_i}), so adjacent sub-images share all low-frequency content by Eq. 5, and the high correlations in Table 1 may partly reflect that shared content rather than predictability of new detail. That is a genuine methodological weakness in the proxy, but it is not circularity in the derivation chain: the paper's headline results are measured by external FID statistics after actual training, not obtained from this correlation, and the ablation study provides independent empirical support for each design choice. The Limitations section notes scalability is unexplored, which is a scope limitation, not a circular step. Verdict: no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The model rests on two domain assumptions about natural image statistics and one design-specific assumption about causal masking. The only hand-set free parameters are the spectral cutoff schedule and the token count. No new physical or conceptual entities are introduced.

free parameters (2)
  • Non-uniform token-frequency mapping schedule (omega_i) = omega_i = i for i in (0,32]; 2i-32 for (32,48]; 12i-512 for (48,64]
    Hand-designed piecewise linear schedule that allocates more tokens to low frequencies; it is a free choice not learned or derived, and it directly affects reconstruction fidelity and token allocation.
  • Number of spectral tokens N = 64 for main experiments, 16x4 for the p4 variant
    Set as a hyperparameter; the paper motivates 64 from token efficiency but does not provide a principled selection procedure.
assumptions (3)
  • domain assumption DCT coefficient magnitudes of natural images follow a power-law distribution concentrated in low frequencies
    Used in Section 3.1 to justify coarse quantization of high frequencies and non-uniform token mapping; cited to image compression literature and power-law references.
  • domain assumption Low-to-high frequency order is a causally meaningful sequence for autoregressive image modeling
    The core premise of the method, introduced in Section 3.1 and tested only with a linear correlation proxy in Section 4.2; not proven.
  • ad hoc to paper Causal masking of spectral tokens in encoder and decoder prevents harmful information leakage without losing reconstruction ability
    The paper claims this in Section 3.2, but the mask also removes bidirectional context that ordinary tokenizers use, and its benefits are only shown in the ablation (Table 4) on one dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SpectralAR: Spectral Autoregressive Visual Generation." pith.science (2026). https://pith.science/paper/KOK7SOFP

@misc{pith2026250610962,
  author       = {Pith},
  title        = {Pith review of: SpectralAR: Spectral Autoregressive Visual Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KOK7SOFP}},
  note         = {Machine review of arXiv:2506.10962}
}
read the original abstract

Autoregressive visual generation has garnered increasing attention due to its scalability and compatibility with other modalities compared with diffusion models. Most existing methods construct visual sequences as spatial patches for autoregressive generation. However, image patches are inherently parallel, contradicting the causal nature of autoregressive modeling. To address this, we propose a Spectral AutoRegressive (SpectralAR) visual generation framework, which realizes causality for visual sequences from the spectral perspective. Specifically, we first transform an image into ordered spectral tokens with Nested Spectral Tokenization, representing lower to higher frequency components. We then perform autoregressive generation in a coarse-to-fine manner with the sequences of spectral tokens. By considering different levels of detail in images, our SpectralAR achieves both sequence causality and token efficiency without bells and whistles. We conduct extensive experiments on ImageNet-1K for image reconstruction and autoregressive generation, and SpectralAR achieves 3.02 gFID with only 64 tokens and 310M parameters. Project page: https://huang-yh.github.io/spectralar/.

Figures

Figures reproduced from arXiv: 2506.10962 by the authors.

Figure 1
Figure 1. We approach autoregressive visual generation from the spectral perspective and propose [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison between autoregressive visual genera￾tion methods. SpectralAR achieves both token efficiency and se￾quence causality with nested 1D spectral tokens. quential order of images, motivating us to represent images as causal spectral sequences. Specifically, we first transform images into spectral tokens with Nested Spectral Tokeniza￾tion (NST), which uses varying sequence lengths to repre￾sent images across di… view at source ↗
Figure 3
Figure 3. Overall pipeline of the proposed SpectralAR. Left: We convert an input image into a 1D causal sequence efficiently with nested spectral tokenization. Each nested sequence is supervised with distinct reconstruction target in a coarse-to-fine manner, which endows each token with an explicit spectral interpretation. We also apply the causal mask to the spectral tokens in the encoder and decoder to enhance the causality… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Frequency of each element having the highest cor￾relation with another one. It represents the source of reference information for predicting the next token. The raster-scan method exhibits excessive dependency on boundaries of image patches. enabling precise modeling o…
Figure 5
Figure 5. Figure 5: Visualizations of generated 256×256 samples on ImageNet-1K. These samples cover a wide range of categories and styles, demonstrating the ability of SpectralAR to generate both diverse and high-quality images. use the ViT-B [9] as the encoder and decoder, and set the nu…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Adaptive Visual Autoregressive Acceleration via Dual-Linkage Entropy Analysis

    cs.CV 2026-02 conditional novelty 5.0 of 10

    A training-free entropy-guided token-pruning framework accelerates VAR image generation up to 2.9× with negligible benchmark loss by activating pruning at an adaptive entropy-growth inflection point and adjusting rati...

Reference graph

Works this paper leans on

71 extracted references · 46 canonical work pages · cited by 1 Pith paper

  1. [11]

    Spectral image tokenizer.arXiv preprint arXiv:2412.09607,

    Carlos Esteves, Mohammed Suhail, and Ameesh Makadia. Spectral image tokenizer.arXiv preprint arXiv:2412.09607,

  2. [43]

    Cart: Compositional auto-regressive transformer for image generation.arXiv preprint arXiv:2411.10180, 2024

    Siddharth Roheda. Cart: Compositional auto-regressive transformer for image generation.arXiv preprint arXiv:2411.10180, 2024. 3

  3. [1]

    Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 2006

    Nasir Ahmed, T Natarajan, and Kamisetty R Rao. Discrete cosine transform.IEEE transactions on Computers, 100(1): 90–93, 2006. 3, 4

  4. [2]

    Courier Corporation, 2003

    Jean Baptiste Joseph Baron Fourier et al.The analytical the- ory of heat. Courier Corporation, 2003. 2, 3

  5. [3]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6

  6. [4]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315–11325, 2022. 2, 6

  7. [5]

    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. InICML, pages 1691–1703. PMLR,

  8. [6]

    Power-law distributions in empirical data.SIAM re- view, 51(4):661–703, 2009

    Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ New- man. Power-law distributions in empirical data.SIAM re- view, 51(4):661–703, 2009. 2

Show all 71 references
  1. [7]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. InCVPR, pages 248–255. Ieee, 2009. 6, 8

  2. [8]

    Diffusion models beat gans on image synthesis.NIPS, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.NIPS, 34:8780–8794, 2021. 6

  3. [9]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint ar...

  4. [10]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InCVPR, pages 12873–12883, 2021. 2, 3, 4, 5, 6, 7

  5. [12]

    Leveraging fre- quency analysis for deep fake image recognition

    Joel Frank, Thorsten Eisenhofer, Lea Sch ¨onherr, Asja Fis- cher, Dorothea Kolossa, and Thorsten Holz. Leveraging fre- quency analysis for deep fake image recognition. InICML, pages 3247–3258. PMLR, 2020. 3

  6. [13]

    Frequency separation for real-world super-resolution

    Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. InICCVW, pages 3599–3608. IEEE, 2019. 3

  7. [14]

    Generative adversarial nets.NIPS, 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.NIPS, 27, 2014. 2, 4

  8. [15]

    Im- age enhancement by nonlinear extrapolation in frequency space.TIP, 9(6):1035–1048, 2000

    Hayit Greenspan, Charles H Anderson, and Sofia Akber. Im- age enhancement by nonlinear extrapolation in frequency space.TIP, 9(6):1035–1048, 2000. 3

  9. [16]

    Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis.arXiv preprint arXiv:2412.04431, 2024

    Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis.arXiv preprint arXiv:2412.04431, 2024. 2

  10. [17]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.NIPS, 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.NIPS, 30, 2017. 6

  11. [18]

    Reducing the dimensionality of data with neural networks.science, 313(5786):504–507, 2006

    Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks.science, 313(5786):504–507, 2006. 3

  12. [19]

    Denoising diffu- sion probabilistic models.NIPS, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NIPS, 33:6840–6851, 2020. 1

  13. [20]

    Cascaded diffusion models for high fidelity image generation.JMLR, 23(47):1– 33, 2022

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation.JMLR, 23(47):1– 33, 2022. 6

  14. [21]

    Owl- 1: Omni world model for consistent long video generation

    Yuanhui Huang, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Di Zhang, Jie Zhou, and Jiwen Lu. Owl- 1: Omni world model for consistent long video generation. arXiv preprint arXiv:2412.09600, 2024. 2

  15. [22]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 4

  16. [23]

    Scaling up gans for text-to-image synthesis

    Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. InCVPR, pages 10124– 10134, 2023. 6

  17. [24]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 3

  18. [25]

    Videopoet: A large language model for zero-shot video gen- eration.arXiv preprint arXiv:2312.14125, 2023

    Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jos ´e Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vigh- nesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video gen- eration.arXiv preprint arXiv:2312.14125, 2023. 5

  19. [26]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7

  20. [27]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.IJCV, 128(7):1956–1981, 2020

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Ui- jlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection...

  21. [28]

    Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recog- nition.Proceedings of the IEEE, 86(11):2278–2324, 1998. 2 9

  22. [29]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. InCVPR, pages 11523–11532, 2022. 2, 3, 6

  23. [30]

    Self- conditioned image generation via generating representations

    Tianhong Li, Dina Katabi, and Kaiming He. Self- conditioned image generation via generating representations. CoRR, 2023. 6

  24. [31]

    Autoregressive image generation without vec- tor quantization.NIPS, 37:56424–56445, 2025

    Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vec- tor quantization.NIPS, 37:56424–56445, 2025. 2

  25. [32]

    Frequency do- main model augmentation for adversarial attack

    Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi- anglong Liu, Jian Zhang, and Jingkuan Song. Frequency do- main model augmentation for adversarial attack. InECCV, pages 549–566. Springer, 2022. 3

  26. [33]

    Frequency-driven imperceptible adversarial attack on semantic similarity

    Cheng Luo, Qinliang Lin, Weicheng Xie, Bizhu Wu, Jin- heng Xie, and Linlin Shen. Frequency-driven imperceptible adversarial attack on semantic similarity. InCVPR, pages 15315–15324, 2022. 3

  27. [34]

    Open-magvit2: An open-source project toward democratizing auto-regressive visual gener- ation.arXiv preprint arXiv:2409.04410, 2024

    Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual gener- ation.arXiv preprint arXiv:2409.04410, 2024. 2

  28. [35]

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

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Explor- ing flow and diffusion-based generative models with scalable interpolant transformers. InECCV, pages 23–40. Springer,

  29. [36]

    Star: Scale-wise text-to- image generation via auto-regressive representations.arXiv preprint arXiv:2406.10797, 2024

    Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to- image generation via auto-regressive representations.arXiv preprint arXiv:2406.10797, 2024. 2

  30. [37]

    Finite scalar quantization: Vq-vae made simple.arXiv preprint arXiv:2309.15505, 2023

    Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple.arXiv preprint arXiv:2309.15505, 2023. 3

  31. [38]

    Correlation and the coefficient of determina- tion.Psychological bulletin, 97(2):307, 1985

    Daniel J Ozer. Correlation and the coefficient of determina- tion.Psychological bulletin, 97(2):307, 1985. 7

  32. [39]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, pages 4195–4205, 2023. 1, 2, 3, 6

  33. [40]

    Improving language understanding by gen- erative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by gen- erative pre-training. 2018. 2, 7

  34. [41]

    Global filter networks for image classification

    Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. Global filter networks for image classification. Advances in neural information processing systems, 34:980– 993, 2021. 3

  35. [42]

    Gener- ating diverse high-fidelity images with vq-vae-2.NIPS, 32,

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Gener- ating diverse high-fidelity images with vq-vae-2.NIPS, 32,

  36. [44]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, pages 10684– 10695, 2022. 1, 3, 6

  37. [45]

    Stylegan- xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan- xl: Scaling stylegan to large diverse datasets. InACM SIG- GRAPH 2022 conference proceedings, pages 1–10, 2022. 6

  38. [46]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1

  39. [48]

    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

  40. [49]

    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. 2, 3, 5

  41. [50]

    Visual autoregressive modeling: Scalable image gen- eration via next-scale prediction.NIPS, 37:84839–84865,

    Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image gen- eration via next-scale prediction.NIPS, 37:84839–84865,

  42. [51]

    Conditional image gen- eration with pixelcnn decoders.NIPS, 29, 2016

    Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image gen- eration with pixelcnn decoders.NIPS, 29, 2016. 3, 5

  43. [52]

    Neural discrete representation learning.NIPS, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.NIPS, 30, 2017. 3

  44. [53]

    Attention is all you need.NIPS, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.NIPS, 30, 2017. 2

  45. [54]

    Extracting and composing robust features with denoising autoencoders

    Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. InICML, pages 1096– 1103, 2008. 3

  46. [55]

    The jpeg still picture compression stan- dard.Communications of the ACM, 34(4):30–44, 1991

    Gregory K Wallace. The jpeg still picture compression stan- dard.Communications of the ACM, 34(4):30–44, 1991. 2, 3, 4, 6

  47. [56]

    Emu3: Next-token prediction is all you need.arXiv preprint arXiv:2409.18869, 2024

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need.arXiv preprint arXiv:2409.18869, 2024. 2, 5

  48. [57]

    Loong: Generating minute-level long videos with autoregressive lan- guage models.arXiv preprint arXiv:2410.02757, 2024

    Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive lan- guage models.arXiv preprint arXiv:2410.02757, 2024. 2

  49. [58]

    Overview of the h

    Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding stan- dard.TCSVT, 13(7):560–576, 2003. 2

  50. [59]

    Janus: Decoupling visual encoding for unified multimodal understanding and generation.arXiv preprint arXiv:2410.13848, 2024

    Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation.arXiv preprint arXiv:2410.13848, 2024. 2

  51. [60]

    Show-o: One single transformer to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528, 2024

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528, 2024. 2 10

  52. [61]

    Wavegan: Frequency-aware gan for high-fidelity few-shot image generation

    Mengping Yang, Zhe Wang, Ziqiu Chi, and Wenyi Feng. Wavegan: Frequency-aware gan for high-fidelity few-shot image generation. InECCV, pages 1–17. Springer, 2022. 3

  53. [63]

    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. 3, 6

  54. [64]

    Scaling autoregres- sive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gun- jan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin- fei Yang, Burcu Karagol Ayan, et al. Scaling autoregres- sive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 3

  55. [65]

    Randomized autoregressive visual generation

    Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang- Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 2

  56. [66]

    An image is worth 32 tokens for reconstruction and generation.NIPS, 37: 128940–128966, 2025

    Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation.NIPS, 37: 128940–128966, 2025. 2, 3, 4, 6, 8

  57. [67]

    Wavefill: A wavelet-based generation network for image inpainting

    Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In ICCV, pages 14114–14123, 2021. 3

  58. [68]

    Cid: Combined image denoising in spatial and frequency do- mains using web images

    Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Cid: Combined image denoising in spatial and frequency do- mains using web images. InCVPR, pages 2933–2940, 2014. 3

  59. [69]

    Var-clip: Text-to-image gen- erator with visual auto-regressive modeling.arXiv preprint arXiv:2408.01181, 2024

    Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziy- ong Feng, and Xingyu Ren. Var-clip: Text-to-image gen- erator with visual auto-regressive modeling.arXiv preprint arXiv:2408.01181, 2024. 2

  60. [70]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, pages 586–595,

  61. [71]

    Llama-adapter: Efficient finetuning of language models with zero-init attention.arXiv preprint arXiv:2303.16199, 2023

    Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Ao- jun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient finetuning of language models with zero-init attention.arXiv preprint arXiv:2303.16199, 2023. 6

  62. [72]

    Movq: Modulating quantized vectors for high- fidelity image generation.NIPS, 35:23412–23425, 2022

    Chuanxia Zheng, Tung-Long Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high- fidelity image generation.NIPS, 35:23412–23425, 2022. 3

  63. [73]

    Occworld: Learning a 3d occupancy world model for autonomous driving

    Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. InECCV, pages 55–72. Springer, 2024. 2 11

Pith tools

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