Pith. sign in

REVIEW 2 major objections 4 minor 52 references

Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Discrete tokens should be conditions, not targets: DisCon models p(xc|xd) and reaches 1.38 gFID on ImageNet-256.

desk verdict A promising two-stage AR design with a real SOTA claim, but the key ablation confounds tokenizer and conditioning, so the mechanism needs one clean control experiment before I'd trust the attribution. read the letter →

arxiv 2507.01756 v2 pith:PQYC6T2V submitted 2025-07-02 cs.CV

classification cs.CV
keywords autoregressiveimagegenerationdiscretetokensasconditionscontinuoustokenmodelingconditionalNet-256two-stagevisualautoregression
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

This paper argues that the right role for discrete tokens in autoregressive image generation is not to be predicted as final outputs but to serve as high-level conditions that guide a continuous autoregressive model. It proposes DisCon, a two-stage framework that first uses a pre-trained discrete autoregressive model (RAR-XXL) to generate discrete tokens encoding global structure, then trains a continuous autoregressive model to predict masked continuous latent tokens conditioned on those discrete tokens and already-seen continuous tokens, with a lightweight diffusion head producing each token. The claim is that the factorization p(xc) = sum over xd of p(xc|xd)p(xd) makes continuous modeling substantially easier while avoiding the information loss of quantization. On ImageNet 256x256, DisCon-L reports gFID 1.38 and rFID 0.28, outperforming the discrete SOTA RAR-XXL (1.48) and the continuous MAR (1.78) while using only 16 AR steps instead of 256. If true, this shows discrete and continuous autoregressive modeling are complementary rather than competing, and that conditioning on strong discrete structure can close the gap between AR and diffusion image quality.

What carries the argument

The load-bearing object is the conditional factorization p(x_c) = sum over x_d of p(x_c|x_d) p(x_d), realized by a two-stage pipeline. Stage one is a pre-trained discrete autoregressive model, RAR-XXL, which generates the discrete token sequence x_d at inference; these tokens are high-level structural conditions, not prediction targets. Stage two is a continuous autoregressive transformer, adapted from MAR, which predicts an intermediate latent variable z_i for each masked continuous token, conditioned on all discrete tokens and previously generated continuous tokens, and a lightweight diffusion head converts z_i into the continuous token x_{c,i}. The factorization is what carries the argument: the discrete tokens are meant to pin down which local mode of the data distribution is being synthesized, leaving the continuous stage to model only within-mode variation, which the paper argues is substantially easier than modeling the unconditional continuous distribution.

What would settle it

Train DisCon-L identically but at inference feed it discrete token sequences drawn from a much weaker model, such as RAR-B or a random permutation of RAR-XXL tokens, and measure gFID; if the gFID stays near 1.38, the discrete conditioning is not doing the claimed work. Alternatively, condition the continuous model on ground-truth tokenizer outputs at inference and check whether the gFID drops well below 1.38; a large gap would show the method's success is bounded by discrete-model sampling error rather than by the continuous stage.

Watch

Extended reading notes

Core claim

The paper's central claim is that images are better described as a finite collection of disjoint continuous distributions than as either a purely discrete token space or a single unbounded continuous space, and that autoregressive generation should therefore factor through both: a discrete model picks the mode, a continuous model fills in the details. DisCon instantiates this by training a continuous autoregressive transformer to maximize p(z_i|x_d, x_{c,<i}) for masked latent positions and then mapping each z_i through a diffusion head to the continuous token x_{c,i}, with the full discrete sequence x_d given as conditioning. At inference, a pre-trained RAR-XXL supplies x_d; no classifier-free guidance is used in the continuous stage. On ImageNet-256 the model reports 1.38 gFID, 0.28 rFID, and IS 325.1, beating RAR-XXL's 1.48 gFID and MAR-L's 1.78 gFID while using 16 rather than 256 autoregressive steps; ablation shows the gain depends on the strength of the discrete model, with RAR-B and RAR-L giving little or no improvement.

Load-bearing premise

The whole result rests on the premise that a strong pre-trained discrete autoregressive model (RAR-XXL) produces discrete conditioning tokens that are both faithful enough to the tokenizer outputs seen in training and informative enough that the conditional distribution p(xc|xd) is genuinely easier to learn than the unconditional p(xc); the paper gives no formal argument for this simplification, and its own Table 3 shows the benefit essentially disappears when the discrete model is weak.

Editorial extensions

If this is right

  • If DisCon's central claim is right, autoregressive image generation no longer needs to choose between discrete fidelity bottlenecks and continuous optimization difficulty: strong discrete tokens can be reused as conditioning, so future AR systems can inherit improvements in discrete tokenizers and discrete AR models directly.
  • The reduction in AR steps, from 256 for MAR to 16 for DisCon at better quality, means continuous AR models conditioned on discrete structure could become practical for high-resolution or video generation where step count dominates cost.
  • Because the architecture is modular, swapping in better discrete AR models or better continuous tokenizers should translate directly into lower gFID; the paper's Table 3 already shows the trend from RAR-B to RAR-XXL.
  • The framework is LLM-compatible: since both stages are autoregressive, DisCon-style conditioning can be integrated into multimodal LLMs that already emit discrete tokens, using those tokens as conditions for continuous image decoding.
  • Reconstruction fidelity and generation fidelity are no longer traded off: rFID matches the continuous tokenizer's 0.28 while gFID beats purely discrete models, so the method points to a regime where reconstruction and generation improve together.

Reading between the lines

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

  • Extrapolating from Table 3, the method turns the quality of the discrete AR model into a first-class driver of continuous generation quality; a natural test is whether an even stronger discrete model pushes gFID below 1.38 without retraining the continuous stage.
  • The paper does not quantify the distribution shift between training-time tokenizer outputs and inference-time RAR-XXL sampled outputs; an implicit experiment is to condition DisCon at inference on ground-truth tokenizer outputs versus RAR-XXL outputs and measure the gFID difference, isolating how much performance is lost to conditioning noise.
  • The 'finite set of disjoint continuous distributions' view suggests a direct extension to other modalities: discrete semantic units such as text tokens or audio codes could condition continuous generators in the same way, making DisCon a general template for hybrid discrete-continuous autoregressive modeling.
  • Since CFG is applied only to the discrete stage, there is an unexplored interaction: tuning CFG jointly across both stages could yield further gains or reveal that discrete-stage CFG alone is sufficient for the diversity-quality trade-off.
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

2 major / 4 minor

Summary. The paper proposes DisCon, a two-stage autoregressive image synthesis method for ImageNet-256. A pre-trained discrete autoregressive model (RAR-XXL) generates discrete tokens, which are used exclusively as conditioning signals rather than as prediction targets. A MAR-style continuous autoregressive model, equipped with a lightweight diffusion head, then predicts continuous tokens from the VAVAE tokenizer, conditioned on the discrete tokens and on previously generated continuous tokens. Training masks a fraction of continuous tokens and supervises their reconstruction given the full discrete-token condition. The authors report gFID 1.38 and rFID 0.28 for DisCon-L (558M parameters), outperforming existing autoregressive baselines, and show that high-quality generation can be achieved with 16 AR steps rather than 256, yielding a roughly 5x inference speedup over MAR. Ablations examine the discrete-conditioning term, the choice of the pre-trained discrete AR model, the number of AR steps, diffusion temperature, and training epochs.

Significance. If the reported numbers hold, the paper makes a useful and timely conceptual contribution: it reframes discrete tokens as structural priors instead of generation targets, and it provides empirical evidence that this conditioning reduces the difficulty of continuous autoregressive modeling. The headline gFID of 1.38 with a 558M-parameter model and 16 AR steps is competitive with much larger diffusion and autoregressive systems, and the rFID of 0.28 confirms the advantage of avoiding quantization. The controlled comparison in Section 4.3, which keeps the LDM tokenizer fixed and adds discrete conditioning, supports the key qualitative claim that conditioning helps. The main limitations are the dependence of the final result on a strong pre-trained discrete AR model and the lack of a formal statement of when the conditional distribution is simpler; both are acknowledged in part but deserve further elaboration.

major comments (2)
  1. [Section 4.3, Table 2] The isolation of the discrete-conditioning effect is performed with the LDM tokenizer, while the headline system uses the VAVAE tokenizer. The paper does not report the intermediate configuration 'MAR + VAVAE without discrete conditioning'. As a result, the contribution of the VAVAE tokenizer to the final gFID of 1.38 is not separated from the contribution of the proposed discrete-conditioning mechanism. I acknowledge the controlled LDM-based comparison in Section 4.3, but the attribution in Section 4.4 ('incorporating discrete tokens ... leads to superior performance') would be fully supported only by adding the missing VAVAE no-condition control, or by reporting a complete 2x2 ablation crossing tokenizer (LDM vs VAVAE) with discrete conditioning (present vs absent).
  2. [Section 3.3, Eq. (9)] The 'Why DisCon Works' argument states that p(xc|xd) is 'significantly simpler' than p(xc), but Eq. (9) is the law of total probability and is true for any joint distribution; it does not by itself imply a reduction in estimation complexity. Please provide a formal argument for the simplification claim, for example in terms of mutual information, support size, or entropy, or explicitly present it as a heuristic hypothesis that is then validated empirically by the ablations. As written, the theoretical motivation is circular: it assumes the very ease of conditional modeling that the method is meant to establish.
minor comments (4)
  1. [Table 2 caption] The caption should state explicitly that all +Condition rows use the LDM tokenizer (i.e., the VAVAE is replaced) so that a reader comparing Table 2 with Table 1 does not mistake the ablation for a simultaneous change of tokenizer and conditioning.
  2. [Tables 1 and 2] The same 427M-parameter model appears as DisCon-B (gFID 1.41) in Table 1 and as '+Condition 427M' (gFID 1.57) in Table 2; the note 'these results are preliminary' is not specific enough. Please clarify whether the Table 2 rows are trained for fewer than the default 800 epochs, and whether they are the same architecture as the corresponding Table 1 models.
  3. [Table 3 and Section 4.3] The finding that the benefit essentially disappears with RAR-B and RAR-L should be presented as a core limitation of the two-stage design in the main text, not only in a table caption, since the method's value is evidently contingent on the quality of the pre-trained discrete prior.
  4. [Section 4.1 and Section 3.3] Training uses ground-truth discrete tokens from the MaskGIT tokenizer, while inference uses tokens generated by RAR-XXL under classifier-free guidance; this train/inference distribution shift is not discussed. Given the sensitivity to the discrete model shown in Table 3, a brief comment or a small experiment on this mismatch would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: DisCon's gains are measured against external FID and the conditioning model is a separate pre-trained baseline.

full rationale

DisCon's central derivation is p(xc) = Σ p(xc|xd) p(xd) (Eq. 9), which is the law of total probability and therefore not circular; it is a decomposition, not an empirical prediction, and the paper does not claim the identity itself proves the performance gain. The gFID/rFID results are computed on 50k generated images against the ImageNet validation distribution using standard external metrics; no parameter is fitted to the metric and then reported as a prediction. The conditioning tokens come from RAR-XXL, an independently published pre-trained model, and Table 3 explicitly evaluates weaker and stronger conditioning models, so the dependence is disclosed rather than smuggled. The ablation in Table 2 changes both the tokenizer and the conditioning relative to the final model, but the text states an LDM-tokenizer variant was run to isolate discrete conditioning; this is an experimental-design concern (a missing MAR+VAVAE no-condition control) rather than a circular reduction. There are no self-citations by the authors, no imported uniqueness theorem, and no renamed known result that is passed off as a derivation. The appended limitation statement in Supplementary F acknowledges reliance on a diffusion head, which is an efficiency caveat, not a circular step. Therefore no circular step meets the evidence bar.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities. Its free parameters are standard sampling and training hyperparameters, but several are tuned per configuration and not fixed, which contributes to the discrepancy between tables. The axioms are mostly assumptions borrowed from prior work; the key new assumption is that discrete conditions simplify the continuous distribution, which is stated but not proven.

free parameters (4)
  • sampling temperature = not reported (tuned per tokenizer)
    Optimal diffusion-head temperature depends on the continuous tokenizer, shown in Figure 7; chosen by validation.
  • number of autoregressive steps = 16 (Table 1), 32 (Table 2)
    Default set to 16, but the same model is reported with 32 steps in the ablation; affects gFID by about 0.2 points.
  • CFG scale for discrete conditioning = default from RAR, not specified
    The discrete tokens that condition the continuous model are generated with RAR's default classifier-free guidance; this scale is not given.
  • EMA decay = 0.9999
    Used for training; authors note it may lag performance (Figure 8), so the final numbers depend on this choice.
assumptions (5)
  • domain assumption Natural images form a finite set of disjoint continuous distributions.
    Motivates the discrete-as-condition design; asserted in Section 3.1 without empirical characterization of mode separation.
  • domain assumption Discrete tokens capture most essential information, making p(xc|xd) simpler than p(xc).
    Core premise of the method; Section 3.3, but no quantitative measure of 'essential information' or 'simplicity' is given.
  • standard math The autoregressive factorization of the conditional continuous process (Eq 5) is valid.
    Carried over from MAR; assumes conditional independence given previous tokens and latent variable.
  • domain assumption The diffusion head learns p(xc,i|zi) with a standard epsilon-prediction objective.
    Borrowed from MAR; no verification on this tokenizer beyond reported FID.
  • domain assumption RAR-XXL and the MaskGIT tokenizer remain fixed and competent when used as conditions.
    The method inherits the failures of the discrete stage; Table 3 shows performance drops with weaker discrete models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis." pith.science (2026). https://pith.science/paper/PQYC6T2V

@misc{pith2026250701756,
  author       = {Pith},
  title        = {Pith review of: Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQYC6T2V}},
  note         = {Machine review of arXiv:2507.01756}
}
abstract

Recent advances in large language models (LLMs) have spurred interests in encoding images as discrete tokens and leveraging autoregressive (AR) frameworks for visual generation. However, the quantization process in AR-based visual generation models inherently introduces information loss that degrades image fidelity. To mitigate this limitation, recent studies have explored to autoregressively predict continuous tokens. Unlike discrete tokens that reside in a structured and bounded space, continuous representations exist in an unbounded, high-dimensional space, making density estimation more challenging and increasing the risk of generating out-of-distribution artifacts. Based on the above findings, this work introduces DisCon (Discrete-Conditioned Continuous Autoregressive Model), a novel framework that reinterprets discrete tokens as conditional signals rather than generation targets. By modeling the conditional probability of continuous representations conditioned on discrete tokens, DisCon circumvents the optimization challenges of continuous token modeling while avoiding the information loss caused by quantization. DisCon achieves a gFID score of 1.38 on ImageNet 256$\times$256 generation, outperforming state-of-the-art autoregressive approaches by a clear margin. Project page: https://pengzheng0707.github.io/DisCon.

Figures

Figures reproduced from arXiv: 2507.01756 by the authors.

Figure 1
Figure 1. Visual Data Representations. Discrete AR models rep￾resent data as separate categories, simplifying learning but intro￾ducing quantization-induced information loss, leading to higher rFID. In contrast, continuous AR models assume data lies in a continuous space, achieving lower rFID. However, unlike discrete tokens that reside in a structured and bounded space, continuous representations exist in an unbounded, high-… view at source ↗
Figure 2
Figure 2. Discrete vs. Continuous AR Models. Top: Images generated by RAR-XXL [46], the SOTA discrete AR model. Bottom: Images generated by our DisCon-L model, a continuous AR model conditioned on the discrete tokens produced by RAR-XXL. Zoom in for better visualization to observe the significant improvements in generation quality. tokenization [13, 16, 26, 31]. These models directly learn continuous latent representations bu… view at source ↗
Figure 3
Figure 3. Overview of the Proposed DisCon Pipeline. Given an input image, discrete and continuous tokens are first extracted using pre-trained tokenizers, with a certain proportion of the continuous tokens masked. An autoregressive model then predicts the masked tokens, conditioned on both the discrete tokens and the available continuous tokens. During inference, a pre-trained discrete AR model (e.g., RAR-XXL [46]) first gene… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Qualitative Results. Images generated by DisCon-B (left) and DisCon-L (right), demonstrating high-fidelity synthesis. value (1.38). We also provide qualitative comparisons in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: Generation Performance during Training. Our method achieves SOTA per￾formance (around 1.5 gFID) after 200 train￾ing epochs. Note that these results are ob￾tained with a 0.9999 EMA setting, which may cause a slight performance lag. De￾tailed training loss curves in the …
Figure 9
Figure 9. Figure 9: Demonstration of Failure Cases. Top: Images generated by RAR-XXL. Bottom: Images generated by our proposed DisCon-L. Although such issues are common in image synthesis, our method exhibits improved performance. A. Implementation Details In our experiments, the training…
Figure 10
Figure 10. Figure 10: shows the training loss curve for DisCon-L. The loss stabilizes at around 100 epochs, demonstrating the re￾duced optimization complexity achieved by our two-stage approach. The efficient training dynamics underscore the benefits of decoupling the modeling of discrete …
Figure 11
Figure 11. Figure 11: Results conditioned on discrete tokens generated by different AR models. From top to bottom: RAR-XXL, RAR-L, and RAR-B. For each model, the top row shows results generated by the respective RAR model, while the bottom row displays outputs from our DisCon method. Zoom …
Figure 12
Figure 12. Figure 12: Class-Conditioned Generation. This figure showcases images generated by DisCon-L across various classes, demonstrating the high fidelity and diversity achieved by our approach [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 9 canonical work pages

  1. [1]

    Gpt-4 technical report

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

  2. [2]

    Maskgit: Masked generative image transformer

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 11315–11325, 2022. 4, 5

  3. [3]

    Softvq-vae: Efficient 1-dimensional con- tinuous tokenizer

    Hao Chen, Ze Wang, Xiang Li, Ximeng Sun, Fangyi Chen, Jiang Liu, Jindong Wang, Bhiksha Raj, Zicheng Liu, and Emad Barsoum. Softvq-vae: Efficient 1-dimensional con- tinuous tokenizer. arXiv preprint arXiv:2412.10958 , 2024. 3

  4. [4]

    Flow matching in latent space

    Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698,

  5. [5]

    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, pages 248–255. Ieee, 2009. 2

  6. [6]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 2

  7. [7]

    Fluid: Scaling autoregressive text-to-image generative models with continuous tokens

    Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 3

  8. [8]

    Mdtv2: Masked diffusion transformer is a strong image synthesizer

    Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389,

Show all 52 references
  1. [9]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2

  2. [10]

    Rethinking the objectives of vector- quantized tokenizers for image synthesis

    Yuchao Gu, Xintao Wang, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Rethinking the objectives of vector- quantized tokenizers for image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7631–7640, 2024. 3

  3. [11]

    Infinity: Scaling bit- wise autoregressive modeling for high-resolution image syn- thesis

    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. 3

  4. [12]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16000– 16009, 2022. 3

  5. [13]

    Acdit: Interpolating autoregressive con- ditional modeling and diffusion transformer

    Jinyi Hu, Shengding Hu, Yuxuan Song, Yufei Huang, Mingxuan Wang, Hao Zhou, Zhiyuan Liu, Wei-Ying Ma, and Maosong Sun. Acdit: Interpolating autoregressive con- ditional modeling and diffusion transformer. arXiv preprint arXiv:2412.07720, 2024. 2, 3, 5, 6

  6. [14]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 2

  7. [15]

    Autoregressive image generation using residual quantization

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11523–11532, 2022. 3

  8. [16]

    Autoregressive image generation without vec- tor quantization

    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, 2025. 2, 3, 5, 6

  9. [17]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. arXiv preprint arXiv:2210.02747, 2022. 3

  10. [18]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 1, 2, 3

  11. [19]

    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. In European Conference on Com- puter Vision, pages 23–40. Springer, 2024. 5

  12. [20]

    Inference-time scaling for diffu- sion models beyond scaling denoising steps

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu- Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffu- sion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732, 2025. 2

  13. [21]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Ad- vances in neural information processing systems, 35:...

  14. [22]

    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. arXiv preprint arXiv:2412.01827, 2024. 3, 5

  15. [23]

    Scalable diffusion models with transformers

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

  16. [24]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2

  17. [25]

    Tokenflow: Unified image tokenizer for multimodal understanding and generation

    Liao Qu, Huichao Zhang, Yiheng Liu, Xu Wang, Yi Jiang, Yiming Gao, Hu Ye, Daniel K Du, Zehuan Yuan, and Xinglong Wu. Tokenflow: Unified image tokenizer for multimodal understanding and generation. arXiv preprint arXiv:2412.03069, 2024. 1

  18. [26]

    Flowar: Scale-wise autoregres- sive image generation meets flow matching

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Flowar: Scale-wise autoregres- sive image generation meets flow matching. arXiv preprint arXiv:2412.15205, 2024. 2, 3, 5, 6

  19. [27]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 6

  20. [28]

    Taming scalable visual tok- enizer for autoregressive image generation

    Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Taming scalable visual tok- enizer for autoregressive image generation. arXiv preprint arXiv:2412.02692, 2024. 3

  21. [29]

    Llamafu- sion: Adapting pretrained language models for multimodal generation

    Weijia Shi, Xiaochuang Han, Chunting Zhou, Weixin Liang, Xi Victoria Lin, Luke Zettlemoyer, and Lili Yu. Llamafu- sion: Adapting pretrained language models for multimodal generation. arXiv preprint arXiv:2412.15188, 2024. 1

  22. [30]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3, 5

  23. [31]

    Hart: Efficient visual generation with hybrid au- toregressive transformer

    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 au- toregressive transformer. arXiv preprint arXiv:2410.10812,

  24. [32]

    Chameleon: Mixed-modal early-fusion foundation models

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

  25. [33]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3

  26. [34]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    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, 2025. 3, 5

  27. [35]

    Metamorph: Multimodal understanding and generation via instruction tuning

    Shengbang Tong, David Fan, Jiachen Zhu, Yunyang Xiong, Xinlei Chen, Koustuv Sinha, Michael Rabbat, Yann LeCun, Saining Xie, and Zhuang Liu. Metamorph: Multimodal understanding and generation via instruction tuning. arXiv preprint arXiv:2412.14164, 2024. 1

  28. [36]

    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, 3

  29. [37]

    Emu3: Next-token prediction is all you need

    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. 1

  30. [38]

    Parallelized autoregressive visual generation

    Yuqing Wang, Shuhuai Ren, Zhijie Lin, Yujin Han, Haoyuan Guo, Zhenheng Yang, Difan Zou, Jiashi Feng, and Xihui Liu. Parallelized autoregressive visual generation. arXiv preprint arXiv:2412.15119, 2024. 3

  31. [39]

    Maskbit: Embedding-free image generation via bit tokens

    Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiao- hui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 5

  32. [40]

    Janus: Decoupling visual encoding for unified multimodal understanding and generation

    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. 1

  33. [41]

    Vila-u: a unified foundation model inte- grating visual understanding and generation

    Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model inte- grating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024

  34. [42]

    Show-o: One single transformer to unify multimodal understanding and generation

    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. 1

  35. [43]

    Reconstruction vs

    Jingfeng Yao and Xinggang Wang. Reconstruction vs. gener- ation: Taming optimization dilemma in latent diffusion mod- els. arXiv preprint arXiv:2501.01423, 2025. 4, 5, 6

  36. [44]

    Vector-quantized image modeling with improved vqgan

    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. 1, 3

  37. [45]

    Language model beats diffusion–tokenizer is key to visual generation

    Lijun Yu, Jos ´e Lezama, Nitesh B Gundavarapu, Luca Ver- sari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion–tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 3, 5

  38. [46]

    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, 3, 4, 5

  39. [47]

    An image is worth 32 tokens for reconstruction and generation

    Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems , 37:128940– 128966, 2025. 3, 5

  40. [48]

    Representation alignment for generation: Training diffu- sion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffu- sion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 5

  41. [49]

    E-car: Efficient continuous autore- gressive image generation via multistage modeling

    Zhihang Yuan, Yuzhang Shang, Hanling Zhang, Tongcheng Fang, Rui Xie, Bingxin Xu, Yan Yan, Shengen Yan, Guo- hao Dai, and Yu Wang. E-car: Efficient continuous autore- gressive image generation via multistage modeling. arXiv preprint arXiv:2412.14170, 2024. 3

  42. [50]

    Fast training of diffusion models with masked transformers

    Hongkai Zheng, Weili Nie, Arash Vahdat, and Anima Anandkumar. Fast training of diffusion models with masked transformers. arXiv preprint arXiv:2306.09305, 2023. 5

  43. [51]

    Transfusion: Pre- dict the next token and diffuse images with one multi-modal model

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Pre- dict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1

  44. [52]

    Scaling the codebook size of vqgan to 100,000 with a utilization rate of 99%

    Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of 99%. arXiv preprint arXiv:2406.11837, 2024. 3 Supplementary Materials Figure 9. Demonstration of Failure Cases. Top: Images generated by RAR-XXL. Bottom: Ima...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.