Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Marrying Autoregressive Transformer and Diffusion with Multi-Reference Autoregression

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TransDiff is a unified framework that trains an autoregressive Transformer as a semantic encoder and a diffusion model as its decoder, achieving FID 1.42 on class-conditional ImageNet 256x256 via Multi-Reference Autoregression.

desk verdict TransDiff is a promising hybrid of AR and diffusion with strong reported ImageNet numbers, but the paper never specifies how the model learns to generate semantic features from an all-mask input — the load-bearing training detail is missing. read the letter →

arxiv 2506.09482 v3 pith:ET3ONT54 submitted 2025-06-11 cs.CV

Dingcheng Zhen , Qian Qiao , Xu Zheng , Tan Yu , Kangxi Wu , Ziwei Zhang , Siyuan Liu , Shunshun Yin
show 1 more author
Ming Tao
This is my paper
classification cs.CV
keywords autoregressiveimagegenerationdiffusionmodelsmulti-referenceautoregressionflowmatchingsemanticfeatureencodingNet256x256conditionalsynthesis
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 introduces TransDiff, a unified model that uses an autoregressive Transformer to turn a class label and image latents into semantic conditions, then a diffusion model to decode those conditions into images. It claims this is the first end-to-end joint training of the two paradigms, reporting FID 1.61 with one-step autoregression and 1.42 with Multi-Reference Autoregression on ImageNet 256x256. The stated advantage is speed: the one-step variant is 2x faster than autoregressive baselines and 112x faster than diffusion-only models. A companion paradigm, MRAR, generates a complete image per step and references several prior generated images, which the paper says improves diversity and quality.

What carries the argument

The load-bearing object is the Transformer encoder's semantic feature map $c = \mathrm{ART}(\mathrm{input})$ in the continuous image-latent space, which replaces discrete VQ tokens and serves as the conditioning signal for a DiT-based diffusion decoder trained with a rectified-flow objective. The mechanism is the joint loss of Eq. (8), where the autoregressive prediction of the next latent block is scored by the diffusion velocity-matching loss against the noised latent, so the AR Transformer learns to produce conditions that the diffusion decoder can interpret. MRAR extends this by concatenating several previously generated image latents to the input, so the condition $c$ is built from multiple complete images rather than from a single pass over a mask.

What would settle it

Train TransDiff with exactly the loss in Eq. (8) and no null-conditioning or masking objective, then run Algorithm 1 with an all-mask input; if the sampled images are not class-consistent or FID is far above 1.61, the reported one-step inference scheme does not work as described.

Watch

Extended reading notes

Core claim

The central claim is that an autoregressive Transformer, operating on continuous VAE latents instead of discrete VQ tokens, can act as a semantic encoder whose output conditions a diffusion decoder, and that the two can be trained jointly end-to-end with a single loss combining the autoregressive prediction error and the flow-matching diffusion loss. The reported discovery is that this hybrid reaches FID 1.42 and IS 301.2 on class-conditional ImageNet 256x256 when augmented with MRAR, beating the paper's reported diffusion-only and autoregressive-only baselines at comparable parameter counts, while retaining the ability to run in one diffusion step (FID 1.61). The authors also report that fusing semantic features from images of different classes yields images with combined content, which they take as evidence that the features are high-level and interpretable.

Load-bearing premise

The central claim assumes the AR Transformer, which is trained to encode real image latents into diffusion conditions, can at inference produce those conditions from only a class token and mask tokens, and the paper describes no masked or null-conditioning training objective that would teach it to do so.

Editorial extensions

If this is right

  • If TransDiff's claim holds, discrete tokenization is not a necessary component for autoregressive image generation; continuous latents can be decoded by diffusion.
  • The same model can trade quality for speed by reducing diffusion steps, since the 1-step AR variant already beats several multi-step baselines.
  • Feature-level mixing and interpolation between classes becomes possible through the AR encoder's semantic space, suggesting conditioning is not limited to class labels.
  • MRAR shows that feeding complete previous images as references outperforms token-level and scale-level autoregression, pointing to a new design axis for generative autoregressive modeling.
  • The speed advantage implies that diffusion decoders can be used as one-step or few-step generators when conditioned on semantic features, narrowing the gap to real-time synthesis.

Reading between the lines

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

  • An immediate testable extension is to train TransDiff with an explicit masking or null-conditioning objective so the AR Transformer learns to synthesize conditions from an all-mask input; the paper does not describe such an objective, so without it the inference scheme in Algorithm 1 may not produce the reported FID.
  • The same joint AR-encoder and diffusion-decoder scheme could be applied to text-to-image and video generation, using text or motion latents as the semantic condition, since MRAR's reference mechanism is not class-specific.
  • The reported 112x speedup is measured against diffusion-only models with many denoising steps; a fairer comparison would count total FLOPs and quality at matched parameter budgets, which the paper does not fully report.
  • If the AR encoder truly learns class-disentangled semantic features, then feature arithmetic in this space could be used for controllable image editing, similar to latent-space arithmetic in other generative models.
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

3 major / 6 minor

Summary. The paper introduces TransDiff, a unified framework that combines an autoregressive Transformer (ART) encoder with a diffusion decoder for class-conditional image generation. The encoder maps a class token and (at inference) mask tokens or previously generated images into high-level semantic features, which condition the diffusion decoder. The authors also propose Multi-Reference Autoregression (MRAR), which generates images iteratively while conditioning on multiple previously generated images. On ImageNet 256x256, TransDiff-H with MRAR reports FID 1.42, improving over a 1-Step AR variant (FID 1.61) and outperforming several diffusion-only and AR-only baselines. The paper claims orders-of-magnitude faster inference than diffusion-only models and introduces a cosine-similarity-based diversity metric to support the claim that MRAR improves diversity and quality.

Significance. If the method works as described, TransDiff could be a meaningful step toward unifying autoregressive and diffusion paradigms, and MRAR introduces an appealing new autoregressive strategy that conditions on complete images rather than tokens or coarse scales. The reported FID gains and speed advantages are practically relevant. The paper also provides a new diagnostic diversity metric and includes a human evaluation. However, the significance is currently tempered by two major issues: (1) the training objective is not aligned with the inference-time conditioning (mask-only input), and (2) the MRAR fine-tuning uses real reference latents while inference uses generated images, leading to an unaddressed distribution shift. The lack of released code or checkpoints prevents empirical verification. The paper's central claims are plausible but not fully supported by the described methodology.

major comments (3)
  1. [§3.4, Eq. (8) and Algorithm 1] The training loss in Eq. (8) conditions the AR Transformer on real image latents (via ART(x0, ..., x_{n-1})), but at inference (Algorithm 1) the input is Concat([C, Mask]) where Mask is an all-mask token. The paper never specifies a masked-autoencoding objective or a scheduled-sampling strategy that would teach ART to produce meaningful semantic features from an all-mask input. As written, the architecture would fail at inference because the condition c would be computed from an input distribution never seen during training. Please clarify how the mask-only condition is trained (e.g., a separate masked-prediction loss, or a special training phase), or provide code/checkpoints to demonstrate that the reported FID can be reproduced.
  2. [§3.5.2, Eq. (11)-(12) and Algorithm 2] During MRAR fine-tuning, Eq. (11) uses real image latents x_img_i from the same label as reference conditions, while Algorithm 2 at inference feeds previously generated images o_img_i into the model. This train/inference mismatch (a form of exposure bias) is not addressed in the paper—no scheduled sampling, iterative refinement training, or noise injection is described. The claimed improvement from MRAR over 1-Step AR (FID 1.61 to 1.42) may not follow from the described training procedure unless the model is trained on its own generated references. Please specify the exact reference inputs used during fine-tuning and any technique used to bridge the train-inference gap.
  3. [§4.2.2, Table 4 and Appendix D] The paper argues that MRAR improves image quality by increasing feature diversity, as measured by the L1 norm of the cosine similarity matrix. However, this metric is used only as a diagnostic and is never validated as a monotone proxy for FID or human-rated quality. The evidence in Table 4 is a single run per configuration (one value for the diversity measure and FID), with no error bars, multiple seeds, or statistical significance. The correlation shown in Figure 5 is anecdotal. Since the central claim of MRAR's superiority rests on this diversity-quality link, the current evidence is not conclusive. Please provide a more rigorous validation, e.g., multiple seeds, scatter plots across checkpoints, or a comparison of the diversity metric against FID on a hold-out set.
minor comments (6)
  1. [§4.2.1 and Table 1] The table numbering in the text is inconsistent: 'Table 2' in §4.2.1 actually refers to the comprehensive comparison table (Table 1), and subsequent references are off by one. Please correct all cross-references to tables and figures.
  2. [Figure 5 caption] The caption states 'Diversity Metric among generated samples drops' while the main text states that diversity improves; this is confusing. Please clarify the direction of the metric (lower cosine similarity means higher diversity) and align the caption with the text.
  3. [Eq. (8)] The meaning of the index n and the sequence length N in Eq. (8) is unclear for the 1-Step AR case. If N=0 for a single image without references, the sum notation is unusual; please define the indexing clearly, including how multiple references are incorporated in the MRAR loss.
  4. [Abstract and Conclusion] The claim of being 'the first' unified framework is not sufficiently qualified. The discussion of BLIP3-o in §2.2 is brief, and other possible hybrid AR-diffusion approaches may exist; please strengthen the positioning and explicitly state the novelty relative to prior hybrid attempts.
  5. [Table 1 and Appendix F] The 'Time' values in Table 1 lack measurement details (hardware, batch size, number of sampling steps for the Euler-Maruyama solver), and the abstract's '2x faster' claim is not directly derivable from the table (e.g., TransDiff-H 0.4s vs MAR-H 2.4s is a 6x difference). Please specify how the speedup factors were computed and under what conditions.
  6. [Appendix C and GitHub link] The paper states a GitHub repository (https://github.com/TransDiff/TransDiff), but no code or checkpoints are actually released. Providing the code and pretrained models would substantially improve reproducibility and allow reviewers to verify the reported FID values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TransDiff's reported FID numbers are empirical outcomes of a stated joint training objective and standard hyperparameter selection, not consequences of a self-referential definition or self-citation.

full rationale

The paper's derivation chain runs from the joint loss in Eq. (8), where the AR Transformer produces conditioning features c = ART(x0, ..., x_{n-1}) from real training latents and the diffusion decoder is trained to reconstruct the next latent, to the reported ImageNet FID values. This is a well-defined training objective, and the benchmark numbers are the outcome of optimizing it, not a quantity that is defined in terms of itself. MRAR adds a fine-tuning stage described by Eqs. (11)-(12), where the input references are real image latents during training and generated images during inference; the choice of n = 4 references in Fig. 7 is ordinary hyperparameter selection, not a fitted parameter later relabeled as a prediction. The diversity metric in Appendix D is computed from the model's own semantic features and used only as a diagnostic correlation with FID; it is not part of the loss and does not define the generation result. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in by citation. The reviewer-identified weakness is a training/inference conditioning mismatch: inference feeds an all-mask input to ART (Algorithm 1), while Eq. (8) describes conditioning on real image latents, and MRAR inference conditions on generated images while fine-tuning conditions on real references. That is a correctness or reproducibility gap about whether the described objective realizes the reported numbers, but it is not circularity, because the reported results are not equivalent to the method's inputs by construction and no equation reduces to another merely by definition. Under the instruction to flag circularity only when a specific reduction can be exhibited, no circular step is present, so the score is 0.

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

The paper introduces no new physical entities. Its central assumptions are the masked-input training behavior of the AR transformer and the validity of the self-defined diversity metric as a quality proxy; both are untested outside this paper.

free parameters (4)
  • s1 = not reported
    Scaling factor for drift term in Euler-Maruyama (Eq. 18), described as a trade-off control; values are never given in the paper.
  • s2 = not reported
    Scaling factor for diffusion/noise term in Euler-Maruyama (Eq. 18); values not reported.
  • number of references n in MRAR = 4
    Appendix E and Figure 7 select 4 references as optimal based on FID; this is a hyperparameter tuned on the evaluation metric.
  • CFG scale = not reported
    Classifier-free guidance is used in the main results (w/ CFG columns) but the guidance scale is not stated.
assumptions (4)
  • ad hoc to paper The AR Transformer can produce meaningful semantic features from an all-mask input at inference, despite training losses that supervise reconstruction from input latents.
    Section 3.5.1 defines 1-Step AR inference with input = [C, Mask], but the training objective in Eq. 8 uses x_n as input; no masked-autoencoding objective is described.
  • ad hoc to paper The semantic feature diversity metric (L1 norm of cosine similarity matrix) is a valid and monotone proxy for image generation quality.
    Appendix D defines this metric; the paper uses it to motivate MRAR based on correlations shown in Fig. 5, without external validation.
  • domain assumption Rectified-flow ODE can be solved with a single Euler step (and with s1/s2 rescaling) without collapsing the sample distribution.
    Section F introduces s1/s2 and Table 3 reports 1-step sampling for TransDiff-L; this is an empirical assumption about the learned velocity field.
  • domain assumption The diffusion decoder can decode semantic features c from the AR transformer into image latents accurately enough to beat VQ decoders.
    This is the core architectural bet of TransDiff; it is motivated by results but not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Marrying Autoregressive Transformer and Diffusion with Multi-Reference Autoregression." pith.science (2026). https://pith.science/paper/ET3ONT54

@misc{pith2026250609482,
  author       = {Pith},
  title        = {Pith review of: Marrying Autoregressive Transformer and Diffusion with Multi-Reference Autoregression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ET3ONT54}},
  note         = {Machine review of arXiv:2506.09482}
}
read the original abstract

We introduce TransDiff, the first image generation model that marries Autoregressive (AR) Transformer with diffusion models. In this joint modeling framework, TransDiff encodes labels and images into high-level semantic features and employs a diffusion model to estimate the distribution of image samples. On the ImageNet 256x256 benchmark, TransDiff significantly outperforms other image generation models based on standalone AR Transformer or diffusion models. Specifically, TransDiff achieves a Frechet Inception Distance (FID) of 1.61 and an Inception Score (IS) of 293.4, and further provides x2 faster inference latency compared to state-of-the-art methods based on AR Transformer and x112 faster inference compared to diffusion-only models. Furthermore, building on the TransDiff model, we introduce a novel image generation paradigm called Multi-Reference Autoregression (MRAR), which performs autoregressive generation by predicting the next image. MRAR enables the model to reference multiple previously generated images, thereby facilitating the learning of more diverse representations and improving the quality of generated images in subsequent iterations. By applying MRAR, the performance of TransDiff is improved, with the FID reduced from 1.61 to 1.42. We expect TransDiff to open up a new frontier in the field of image generation.

Figures

Figures reproduced from arXiv: 2506.09482 by the authors.

Figure 1
Figure 1. Generated samples from TransDiff trained on ImageNet. Top: 512×512 and 256×256 samples. Middle: effect of semantic feature diversity in TransDiff on image quality(left to right: increasing diversity.). Bottom: results of semantic features fusion. (The first two columns show images from two classes; the third shows the fused result.) Abstract We present TransDiff, the first image generation model that marries Autoreg… view at source ↗
Figure 2
Figure 2. Comparison of Token-Level AR, Scale-Level AR and MRAR paradigms. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of our Transdiff. Wherein (a) and (c) are the Vanilla Autoregressive Transformer [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 1
Figure 1. Figure 1: In the Middle, the results generated by the diffusion decoder using [PITH_FULL_IMAGE:figures/full_fig_p008_1.png]
Figure 4
Figure 4. Figure 4: Results of semantic features fusion from images of different classes. (The first two columns [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The FID score consistently decreases while Diversity Metric among generated samples drops, indicating improved image quality and diversity. 4.2.3 MRAR beats Token AR and Scale-Level AR paradigms Based on the experimental analysis in AR Transformer’s representation, we …
Figure 6
Figure 6. Figure 6: Human evaluation on our TransDiff and other SoTA methods of Token-Level AR and [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Correlation between the number of references and FID score. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Model comparison on ImageNet 256 × 256 benchmark. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Qualitative Results. More 256×256 class-conditional generation results from TransDiff-H, MRAR on ImageNet. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Qualitative Results. More 512 × 512 class-conditional generation results from TransDiff￾L, MRAR on ImageNet. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 19 canonical work pages

  1. [1]

    The euler scheme for stochastic differential equations: error analysis with malliavin calculus

    Vlad Bally and Denis Talay. The euler scheme for stochastic differential equations: error analysis with malliavin calculus. Mathematics and computers in simulation, 38(1-3):35–41, 1995

    work page 1995

  2. [2]

    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

    arXiv 2018

  3. [3]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  4. [4]

    Efficient-vqgan: Towards high-resolution image generation with efficient vision transformers

    Shiyue Cao, Yueqin Yin, Lianghua Huang, Yu Liu, Xin Zhao, Deli Zhao, and Kaigi Huang. Efficient-vqgan: Towards high-resolution image generation with efficient vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7368–7377, 2023

    work page 2023

  5. [5]

    Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 11305–11315. IEEE, 2022

    work page 2022

  6. [6]

    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 vision and pattern recognition, pages 11315–11325, 2022

    2022

  7. [7]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset. arXiv preprint arXiv:2505.09568, 2025

    arXiv 2025

  8. [8]

    Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91. Springer, 2024

    2024

Show all 50 references
  1. [9]

    Generative pretraining from pixels

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691–1703. PMLR, 2020

    2020

  2. [10]

    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

    2009

  3. [11]

    Diffusion models beat gans on image synthesis

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

    2021

  4. [12]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthes...

    2024

  5. [13]

    Taming transformers for high-resolution image synthesis

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

    2021

  6. [14]

    Diffit: Diffusion vision transformers for image generation

    Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. In European Conference on Computer Vision, pages 37–55. Springer, 2024

    2024

  7. [15]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    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 equilibrium. Advances in neural information processing systems, 30, 2017. 11

    2017

  8. [16]

    Denoising diffusion probabilistic models

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

    2020

  9. [17]

    Fleet, Mohammad Norouzi, and Tim Salimans

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

    2022

  10. [18]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10124–10134, 2023

    2023

  11. [19]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174–24184, 2024

    2024

  12. [20]

    Kloeden and E

    P.E. Kloeden and E. Platen. Numerical Solution of Stochastic Differential Equations. Stochastic Modelling and Applied Probability. Springer Berlin Heidelberg, 2013

    2013

  13. [21]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux , 2024

    2024

  14. [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. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 11513– 11522. IEEE, 2022

    2022

  15. [23]

    Mage: Masked generative encoder to unify representation learning and image synthesis

    Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2142–2152, 2023

    2023

  16. [24]

    Autoregressive image generation without vector quantization

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

    2024

  17. [25]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding

    Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Ji...

    2024 arXiv

  18. [26]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023

    2023

  19. [27]

    Alleviat- ing distortion in image generation via multi-resolution diffusion models and time-dependent layer normalization

    Qihao Liu, Zhanpeng Zeng, Ju He, Qihang Yu, Xiaohui Shen, and Liang-Chieh Chen. Alleviat- ing distortion in image generation via multi-resolution diffusion models and time-dependent layer normalization. Advances in Neural Information Processing Systems, 37:133879–133907, 2024

    2024

  20. [28]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023

    2023

  21. [29]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  22. [30]

    Star: Scale-wise text-conditioned autoregressive image generation

    Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Biye Li, Huaian Chen, and Yi Jin. Star: Scale-wise text-conditioned autoregressive image generation. arXiv preprint arXiv:2406.10797, 2024

    2024 arXiv

  23. [31]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162–8171. PMLR, 2021. 12

    2021

  24. [32]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  25. [33]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172–4182. IEEE, 2023

    2023

  26. [34]

    Language models are unsupervised multitask learners

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

    2019

  27. [35]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

  28. [36]

    Generating diverse high-fidelity images with vq-vae-2

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

    2019

  29. [37]

    Yuille, and Liang-Chieh Chen

    Sucheng Ren, Qihang Yu, Ju He, Xiaohui Shen, Alan L. Yuille, and Liang-Chieh Chen. Beyond next-token: Next-x prediction for autoregressive visual generation. CoRR, abs/2502.20388, 2025

    2025 arXiv

  30. [38]

    High- resolution image synthesis with latent diffusion models

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

    2022

  31. [39]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674–10...

    2022

  32. [40]

    Improved techniques for training gans

    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

    2016

  33. [41]

    Stylegan-xl: Scaling stylegan to large diverse datasets

    Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1–10, 2022

    2022

  34. [42]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-...

    2021

  35. [43]

    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

    2024 arXiv

  36. [44]

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

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

    2024

  37. [45]

    Con- ditional image generation with pixelcnn decoders

    Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Con- ditional image generation with pixelcnn decoders. Advances in neural information processing systems, 29, 2016

    2016

  38. [46]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017

    2017

  39. [47]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  40. [48]

    Language model beats diffusion–tokenizer is key to visual generation

    Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, 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. 13

    arXiv

  41. [49]

    Randomized autore- gressive visual generation

    Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autore- gressive visual generation. CoRR, abs/2411.00776, 2024

    2024 arXiv

  42. [50]

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

    Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling.arXiv preprint arXiv:2408.01181, 2024. 14 A Appendix B Inference Process of 1-Step AR and MRAR The inference process of 1-Step A...

    2024 arXiv

Pith tools

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