Pith. sign in

REVIEW 3 major objections 5 minor 64 references

Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots

T0 review · 3 major / 5 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 This paper claims that masked autoregressive image generation improves when the model first predicts a small set of low-resolution tokens as global-structure pivots and then uses their hidden states to guide dense-token prediction…

desk verdict A solid, incremental extension of MAR with plausible FID gains, but the compute-savings claim and the w/o CFG comparison are under-documented, and the pivot hidden-state extraction needs explicit handling. read the letter →

arxiv 2505.20288 v1 pith:PUTYSQ2F submitted 2025-05-26 cs.CV cs.MM

Guangting Zheng , Yehao Li , Yingwei Pan , Jiajun Deng , Ting Yao , Yanyong Zhang , Tao Mei This is my paper
classification cs.CVcs.MM
keywords hierarchicalautoregressivemodelsmaskedgenerationimagecontinuoustokensdiffusiontransformerheadlow-resolutiontokenpivotsclass-conditionaltext-to-image
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 tries to establish that masked autoregressive image generation improves when the model first predicts a small set of low-resolution tokens that encode global structure, then uses those tokens as pivots to guide prediction of the full-resolution token sequence. On ImageNet 256×256, the proposed Hi-MAR model lowers FID from 2.31 to 1.93 for the base size, improving on the single-scale MAR baseline while using about 54% of its inference compute; on MS-COCO text-to-image generation it reaches FID 4.77, beating comparable masked autoregressive and diffusion baselines. A new Diffusion Transformer head, which attends across all masked and unmasked tokens during denoising, contributes much of the gain. If correct, the work shows that global-to-local hierarchy can be built without retraining a multi-scale tokenizer and without the train/inference mismatch that plagues conditioning on raw low-resolution tokens.

What carries the argument

The load-bearing object is the two-phase hierarchical masked autoregressive Transformer: phase one treats a short sequence of low-resolution continuous tokens as the global-structure pivots, and phase two predicts the dense tokens while attending to the first phase's conditional tokens. The second key piece is the Diffusion Transformer head, a stack of self-attention blocks that denoises all masked tokens jointly using summed time embedding and conditional-token context, so interdependencies among tokens are modeled during prediction. A third component is the scale-aware Transformer block, which injects a resolution-specific scale vector via AdaLN-Zero into every layer so one shared backbone can handle both scales without ambiguity.

What would settle it

Run Hi-MAR's second phase with the first phase fed ground-truth low-resolution images instead of predicted ones, and compare FID to the full pipeline; a large gap would indicate that the hidden-state distribution still differs between training and inference, while a small gap would confirm the conditional-token pivots resolve the discrepancy the paper targets.

Watch

Extended reading notes

Core claim

The central claim is that modeling an image as a two-level token hierarchy, first a handful of continuous low-resolution tokens and then the dense high-resolution tokens, lets a masked autoregressive model capture global context early, where single-scale next-token prediction struggles. Hi-MAR runs two masked autoregressive phases: phase one reconstructs 128×128 tokens under a high masking ratio, and the hidden states of the Hi-MAR Transformer at that scale, not the reconstructed visual tokens, become the pivots for phase two. Conditioning on these conditional tokens rather than on raw low-resolution tokens is what removes the standard training-inference discrepancy of next-scale models. In phase two, a Diffusion Transformer head replaces the per-token MLP head of MAR so that denoising each masked token can see all other tokens, and a learnable scale vector injected through AdaLN-Zero makes the shared backbone resolution-aware. The reported results are FID 1.93, 1.66, and 1.52 for Base, Large, and Huge on ImageNet with classifier-free guidance, versus 2.31, 1.78, and 1.55 for MAR, and FID 4.77 on MS-COCO.

Load-bearing premise

The second phase is conditioned on hidden-state pivots from the first phase, and the paper assumes those hidden states at inference, when the first phase sees fully predicted low-resolution tokens, match the distribution they had during training, when the first phase saw randomly masked low-resolution tokens.

Editorial extensions

If this is right

  • The base-size Hi-MAR reaches FID 1.93 on ImageNet with classifier-free guidance, an absolute improvement of 0.38 over MAR-B, while the paper reports only 54% of MAR's computational cost at similar quality.
  • Large and Huge variants continue the trend with FID 1.66 and 1.52, suggesting the hierarchy helps as model scale grows.
  • On MS-COCO text-to-image generation, Hi-MAR-S achieves FID 4.77, beating AutoNAT-S by 0.59 and MAR by 1.59, and improves several T2I-CompBench composition scores.
  • The second phase saturates at roughly 4 autoregressive steps when the first phase is given 32 steps, meaning the global pivots let the dense stage spend its computation on local detail rather than rediscovering structure.
  • The ablation attributes most of the gain to conditioning on conditional tokens (FID 2.31 to 2.07) and to the Diffusion Transformer head (FID 2.07 to 1.98), with the scale vector adding the final improvement to 1.93.

Reading between the lines

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

  • Because the pivots are hidden states rather than reconstructed pixels, the same two-phase recipe could be applied to other dense sequence modalities, such as video frames or long panoramic images, whenever a cheap global sketch can be produced first.
  • The 54% compute figure is measured for the base model at a fixed batch size on one H100; the practical speedup at other batch sizes or with a smaller first-phase model is not reported and could be tested directly.
  • A direct test of the central assumption would be to feed ground-truth low-resolution tokens at inference in place of the predicted ones: the FID gap between that oracle and the full pipeline would quantify the remaining train/inference mismatch.
  • If the hierarchy is what carries the gain, then a three-level cascade with an even coarser root (for example, 64×64 pivots before 128×128) should further improve early-token global context, though it would also add a phase; this is not explored in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Hi-MAR, a hierarchical masked autoregressive model for image generation. Hi-MAR adds a first phase that operates on a small number of low-resolution visual tokens, whose Transformer hidden states (conditional tokens Z^s) are then used as pivots to condition a second phase that predicts dense tokens. The paper also replaces the MLP-based diffusion head in the second phase with a Diffusion Transformer head, and inserts learnable scale vectors via adaLN-Zero to distinguish the two resolution phases. Experiments on ImageNet 256x256 and MS-COCO/T2I-CompBench report improvements over MAR and other autoregressive and diffusion baselines; the headline numbers are an FID of 1.93 for Hi-MAR-B versus 2.31 for MAR-B under CFG, and a claimed 54% computational cost relative to MAR. The central claim is that a coarse-to-fine, hierarchical pivot structure supplies global context and improves both generation quality and inference efficiency.

Significance. If the results hold, the paper makes a useful empirical contribution: it gives a concrete coarse-to-fine design for masked autoregressive image generation, demonstrates that conditioning on hidden states rather than decoded low-resolution tokens helps, and shows that a transformer-based diffusion head can improve token prediction. The with-CFG ImageNet results are the strongest evidence, and the ablation in Table 5 provides a clean decomposition of the three proposed components. The code release is also a practical strength. However, the headline efficiency claim is not backed by any measurement, the train/inference consistency argument is incomplete, and the w/o CFG comparison is not protocol-matched, so the paper needs additional evidence before the central claims can be accepted.

major comments (3)
  1. [Section 3.2 and 4.2] The claim that conditioning the second phase on Z^s (the Transformer outputs for the low-resolution tokens) removes the training/inference discrepancy is not supported. During training, the first-phase masking ratio is sampled in [0.7, 1.0] (Section 4.2), so the low-resolution sequence always contains ground-truth tokens. At inference, after 32 autoregressive steps, the low-resolution sequence is fully predicted, yet the manuscript does not state how Z^s is computed at that point (e.g., a final forward pass on the fully predicted sequence, or the hidden states from the last autoregressive step), nor does it show that hidden states from fully generated inputs match hidden states from masked ground-truth inputs. Table 5, row 3 only compares conditioning on Z^s against conditioning on decoded visual tokens under a single protocol; it does not quantify the residual mismatch. Please specify the exact inference-time extraction of Z^s and add an experiment that isolates the mismatch, such as comparing second-phase FID when Z^s is computed from ground-truth low-resolution tokens versus from fully generated low-resolution tokens.
  2. [Abstract and Figure 3] The abstract and introduction state that Hi-MAR requires only 54% of the computational costs of MAR, but the paper reports no measurement that produces this number. Figure 3 is a speed/accuracy plot, but it does not define the compute metric (latency, FLOPs, or total step count), does not report numeric values, and does not state which MAR operating point is the baseline. The authors should provide the measurement protocol (hardware, batch size, step counts for both phases, and whether the 54% refers to wall-clock time, FLOPs, or something else) and verify the claim with a table instead of relying on an unextracted curve.
  3. [Section 4.3, Table 2] The w/o CFG rows for Hi-MAR are not comparable to the w/o CFG rows for the baselines. The text states that 'the CFG is only turned off during the prediction of dense tokens' for Hi-MAR, meaning the first phase still uses CFG, whereas MAR and the other autoregressive methods are evaluated with CFG completely disabled. This gives Hi-MAR an unfair advantage in those columns and makes the unqualified statement in the abstract that Hi-MAR outperforms typical AR baselines too broad. The authors should either relabel these rows (e.g., 'CFG in phase one only'), apply the same partially guided protocol to the baselines, or remove the w/o CFG rows and rely on the with-CFG comparison, which is the clean evidence.
minor comments (5)
  1. [Section 3.2] The notation Z^s = {z^s_1, ..., z^s_N} reuses the symbol N already defined as the high-resolution token count in Section 3.1; please use N_s for the low-resolution sequence length.
  2. [Equation (3)] In the Diffusion Transformer head, the context vector c is described as the sum of the time-step embedding and the conditional tokens, but the block equation does not show how c is projected or whether it depends on the block index; please clarify the exact combination and projection.
  3. [Table 5] Replacing the MLP-based Diff. Head2 with the Diffusion Transformer head decreases the parameter count from 245M (row 3) to 239M (row 4), despite the transformer head containing additional self-attention layers; please explain the parameter accounting or correct the table.
  4. [Table 4] The Spatial score for U-ViT-S/2 (Deep) is reported as 0.0353, which is an order of magnitude lower than the other entries in the same row and looks like a typo; please verify this value.
  5. [Related Work and Figure 4] Please fix the typo 'VQV AE' in the related work and make the citation of 'MUSE' consistent; the Figure 4 caption also uses 'low-resolution/high-resolution' while the text refers to first/second phases, so the terminology should be aligned.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Hi-MAR is an empirical architecture evaluated against external FID benchmarks, and no claimed prediction reduces by construction to fitted inputs or to a self-citation chain.

full rationale

The paper's central claims are empirical: Hi-MAR achieves a 0.38 FID improvement over MAR with 54% of the computational cost, measured on ImageNet and MS-COCO against external baselines. These outcomes are not derived from fitted constants; the model is trained with standard masked-token and diffusion losses, and the reported numbers are benchmark evaluations. The low-resolution pivot design (Section 3.2) is an architectural choice, and the paper explicitly compares conditioning on visual tokens versus conditional tokens in the ablation (Table 5). No equation in the paper defines the predicted quantity in terms of a fitted parameter, and no claim reduces to a self-citation. The skeptical concern about train/inference discrepancy in the extraction of conditional tokens Z^s is a validity and robustness issue, not a circularity issue: the paper does not define Z^s as the quantity it later predicts. Self-citations to prior works by overlapping authors appear in the references, but they are not load-bearing for the main result; the key comparison is against MAR (Li et al., 2024), an external method. There is no imported uniqueness theorem, no ansatz smuggled via citation, and no renaming of a known empirical pattern presented as a derivation. The analysis is therefore self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The central claim relies on empirical design choices rather than mathematical derivation; the main load-bearing assumptions are the representational separation of global and local information and the distributional consistency of the learned pivots.

free parameters (4)
  • low-resolution phase scale = 128x128 image tokens for a 256x256 output
    Chosen by hand as the single coarse scale; no search over alternative scales is reported.
  • autoregressive inference steps = 32 steps in phase one, 4 steps in phase two
    Selected from the speed/accuracy curves in Figure 4 for the main results.
  • masking ratio schedule = Beta(4,1) for COCO; [0.7,1.0] for ImageNet phase one; cosine for phase two
    Borrowed from AutoNAT and MAR/MaskGIT; affects training and is not derived.
  • diffusion head depth = 6/8/12 blocks for Base/Large/Huge
    Architectural capacity choices that trade quality against compute.
assumptions (4)
  • domain assumption Low-resolution tokens encode global image structure and high-resolution tokens encode local detail.
    Motivates the two-phase hierarchy in Section 3.2 and Figure 1(c); treated as given rather than demonstrated.
  • domain assumption Transformer hidden states from the first phase remain distributionally consistent between training (masked low-res inputs) and inference (predicted low-res inputs).
    Central to the claimed discrepancy fix in Section 3.2; the paper provides no empirical distribution analysis.
  • domain assumption The pre-trained KL-16 VAE from MAR can encode both 128x128 and 256x256 images without retraining for the new multi-scale use.
    Section 4.2 reuses the MAR VAE for both phases; no validation of the low-resolution latents is given.
  • standard math The standard diffusion denoising loss (Eq. 1) and masked-token prediction objectives are valid training signals.
    Inherited from MAR and diffusion literature; not re-derived here.
invented entities (3)
  • Scale vector v
    purpose: Per-resolution sinusoidal embedding injected through adaLN-Zero to tell the transformer which scale it is processing.
    An internal conditioning vector; its benefit is visible only in the last ablation row (FID 1.98 to 1.93).
  • Conditional token pivots Z_s
    purpose: First-phase transformer hidden states used as global guidance for second-phase dense token prediction.
    An internal representation with no observable handle outside the model; the paper's key new mechanism.
  • Diffusion Transformer head
    purpose: Replaces the per-token MLP diffusion head with a small self-attention diffusion model over all tokens in phase two.
    A new head architecture; contributed gain is shown in the fourth ablation row (FID 2.07 to 1.98).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots." pith.science (2026). https://pith.science/paper/PUTYSQ2F

@misc{pith2026250520288,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PUTYSQ2F}},
  note         = {Machine review of arXiv:2505.20288}
}
read the original abstract

Autoregressive models have emerged as a powerful generative paradigm for visual generation. The current de-facto standard of next token prediction commonly operates over a single-scale sequence of dense image tokens, and is incapable of utilizing global context especially for early tokens prediction. In this paper, we introduce a new autoregressive design to model a hierarchy from a few low-resolution image tokens to the typical dense image tokens, and delve into a thorough hierarchical dependency across multi-scale image tokens. Technically, we present a Hierarchical Masked Autoregressive models (Hi-MAR) that pivot on low-resolution image tokens to trigger hierarchical autoregressive modeling in a multi-phase manner. Hi-MAR learns to predict a few image tokens in low resolution, functioning as intermediary pivots to reflect global structure, in the first phase. Such pivots act as the additional guidance to strengthen the next autoregressive modeling phase by shaping global structural awareness of typical dense image tokens. A new Diffusion Transformer head is further devised to amplify the global context among all tokens for mask token prediction. Extensive evaluations on both class-conditional and text-to-image generation tasks demonstrate that Hi-MAR outperforms typical AR baselines, while requiring fewer computational costs. Code is available at https://github.com/HiDream-ai/himar.

Figures

Figures reproduced from arXiv: 2505.20288 by the authors.

Figure 1
Figure 1. a) Next-token autoregressive (AR). GPT-style autoregressive models process 2D image tokens as a 1D sequence, predicting tokens in raster order using causal attention to ensure each token depends only on preceding ones. b) Next-token masked autoregressive model (MAR). BERT-style autoregressive models initially consider all tokens to be masked, subsequently predicting each masked token based on the known tokens in a r… view at source ↗
Figure 2
Figure 2. (a) Pipeline of conventional hierarchical MAR. Conventional hierarchical mar uses a shared Transformer for both phases and directly leverages low-resolution visual tokens to guide second-phase predictions. (b) Pipeline of Hi-MAR. An image and its low-resolution counterpart are converted into token sequences at two scales. During inference, both sequences are initially masked. In the first phase, masked low-resolutio… view at source ↗
Figure 3
Figure 3. Speed/accuracy trade-off. (a) low-resolution (b) high-resolution [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of autoregressive steps. The experiments are conducted using the Hi-MAR-B model. For experiments varying low-resolution inference steps, typical dense inference steps are fixed at 4. Similarly, for experiments varying typical dense infer￾ence steps, low-resoluti…
Figure 5
Figure 5. Figure 5: Qualitative results on class-conditional image generation and text-to-image generation. The top rows show class-conditional generation, while the bottom rows show text-to-image generation [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 29 canonical work pages

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Qwen technical report

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  3. [3]

    All are worth words: A vit backbone for diffusion models

    Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., and Zhu, J. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 22669--22679, 2023

    work page 2023

  4. [4]

    Large scale GAN training for high fidelity natural image synthesis

    Brock, A., Donahue, J., and Simonyan, K. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=B1xsqj09Fm

    2019

  5. [5]

    D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33: 0 1877--1901, 2020

    1901

  6. [6]

    Chang, H., Zhang, H., Jiang, L., Liu, C., and Freeman, W. T. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11315--11325, 2022

    2022

  7. [7]

    T., Rubinstein, M., et al

    Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K., Freeman, W. T., Rubinstein, M., et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pp.\ 4055--4075, 2023

    work page 2023

  8. [8]

    Controlstyle: Text-driven stylized image generation using diffusion priors

    Chen, J., Pan, Y., Yao, T., and Mei, T. Controlstyle: Text-driven stylized image generation using diffusion priors. In Proceedings of the 31st ACM International Conference on Multimedia, pp.\ 7540--7548, 2023

    work page 2023

Show all 64 references
  1. [9]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 248--255. Ieee, 2009

    2009

  2. [10]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34: 0 8780--8794, 2021

    2021

  3. [11]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  4. [12]

    Taming transformers for high-resolution image synthesis

    Esser, P., Rombach, R., and Ommer, B. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12873--12883, 2021

    2021

  5. [13]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

    2024

  6. [14]

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

    Fan, L., Li, T., Qin, S., Li, Y., Sun, C., Rubinstein, M., Sun, D., He, K., and Tian, Y. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024

    2024 arXiv

  7. [15]

    Fan, W.-C., Chen, Y.-C., Chen, D., Cheng, Y., Yuan, L., and Wang, Y.-C. F. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp.\ 579--587, 2023

    2023

  8. [16]

    Vector quantized diffusion model for text-to-image synthesis

    Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., and Guo, B. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10696--10706, 2022

    2022

  9. [17]

    Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis

    He, W., Fu, S., Liu, M., Wang, X., Xiao, W., Shu, F., Wang, Y., Zhang, L., Yu, Z., Li, H., et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 17123--17131, 2025

    2025

  10. [18]

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

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017

    2017

  11. [19]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    2021

  12. [20]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  13. [21]

    J., Norouzi, M., and Salimans, T

    Ho, J., Saharia, C., Chan, W., Fleet, D. J., Norouzi, M., and Salimans, T. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23 0 (47): 0 1--33, 2022

    2022

  14. [22]

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation

    Huang, K., Sun, K., Xie, E., Li, Z., and Liu, X. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36: 0 78723--78747, 2023

    2023

  15. [23]

    Scaling up gans for text-to-image synthesis

    Kang, M., Zhu, J.-Y., Zhang, R., Park, J., Shechtman, E., Paris, S., and Park, T. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10124--10134, 2023

    2023

  16. [24]

    Kenton, J. D. M.-W. C. and Toutanova, L. K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, volume 1. Minneapolis, Minnesota, 2019

    2019

  17. [25]

    Improved precision and recall metric for assessing generative models

    Kynk \"a \"a nniemi, T., Karras, T., Laine, S., Lehtinen, J., and Aila, T. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019

    2019

  18. [26]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning, pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

    2000

  19. [27]

    Autoregressive image generation using residual quantization

    Lee, D., Kim, C., Kim, S., Cho, M., and Han, W.-S. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11523--11532, 2022

    2022

  20. [28]

    Autoregressive image generation without vector quantization

    Li, T., Tian, Y., Li, H., Deng, M., and He, K. Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems, 37: 0 56424--56445, 2024

    2024

  21. [29]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pp....

    2014

  22. [30]

    Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action

    Lu, J., Clark, C., Lee, S., Zhang, Z., Khosla, S., Marten, R., Hoiem, D., and Kembhavi, A. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 2...

    2024

  23. [31]

    Fit: Flexible vision transformer for diffusion model

    Lu, Z., Wang, Z., Huang, D., Wu, C., Liu, X., Ouyang, W., and Bai, L. Fit: Flexible vision transformer for diffusion model. In International Conference on Machine Learning, pp.\ 33160--33176. PMLR, 2024 b

    2024

  24. [32]

    Revisiting non-autoregressive transformers for efficient image synthesis

    Ni, Z., Wang, Y., Zhou, R., Guo, J., Hu, J., Liu, Z., Song, S., Yao, Y., and Huang, G. Revisiting non-autoregressive transformers for efficient image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7007--7016, 2024

    2024

  25. [33]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4195--4205, 2023

    2023

  26. [34]

    Boosting diffusion models with moving average sampling in frequency domain

    Qian, Y., Cai, Q., Pan, Y., Li, Y., Yao, T., Sun, Q., and Mei, T. Boosting diffusion models with moving average sampling in frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8911--8920, 2024

    2024

  27. [35]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog, 1 0 (8): 0 9, 2019

    2019

  28. [36]

    Flowar: Scale-wise autoregressive image generation meets flow matching

    Ren, S., Yu, Q., He, J., Shen, X., Yuille, A., and Chen, L.-C. Flowar: Scale-wise autoregressive image generation meets flow matching. arXiv preprint arXiv:2412.15205, 2024 a

    2024 arXiv

  29. [37]

    Arvideo: Autoregressive pretraining for self-supervised video representation learning

    Ren, S., Zhu, H., Wei, C., Li, Y., Yuille, A., and Xie, C. Arvideo: Autoregressive pretraining for self-supervised video representation learning. arXiv preprint arXiv:2405.15160, 2024 b

    2024 arXiv

  30. [38]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10684--10695, 2022

    2022

  31. [39]

    Improved techniques for training gans

    Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., and Chen, X. Improved techniques for training gans. Advances in Neural Information Processing Systems, 29, 2016

    2016

  32. [40]

    Stylegan-xl: Scaling stylegan to large diverse datasets

    Sauer, A., Schwarz, K., and Geiger, A. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pp.\ 1--10, 2022

    2022

  33. [41]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019

    2019

  34. [42]

    and Ermon, S

    Song, Y. and Ermon, S. Improved techniques for training score-based generative models. Advances in Neural Information Processing Systems, 33: 0 12438--12448, 2020

    2020

  35. [43]

    Autoregressive model beats diffusion: Llama for scalable image generation

    Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., and Yuan, Z. Autoregressive model beats diffusion: Llama for scalable image generation. CoRR, abs/2406.06525, 2024. URL https://doi.org/10.48550/arXiv.2406.06525

    2024 doi arXiv

  36. [44]

    Hart: Efficient visual generation with hybrid autoregressive transformer

    Tang, H., Wu, Y., Yang, S., Xie, E., Chen, J., Chen, J., Zhang, Z., Cai, H., Lu, Y., and Han, S. Hart: Efficient visual generation with hybrid autoregressive transformer. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  37. [45]

    Df-gan: A simple and effective baseline for text-to-image synthesis

    Tao, M., Tang, H., Wu, F., Jing, X.-Y., Bao, B.-K., and Xu, C. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pp.\ 16515--16525, 2022

    2022

  38. [46]

    Chameleon: Mixed-modal early-fusion foundation models

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

    2024 arXiv

  39. [47]

    M., Hauth, A., Millican, K., et al

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  40. [48]

    G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al

    Team, G., Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  41. [49]

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

    Tian, K., Jiang, Y., Yuan, Z., Peng, B., and Wang, L. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems, 37: 0 84839--84865, 2024

    2024

  42. [50]

    Givt: Generative infinite-vocabulary transformers

    Tschannen, M., Eastwood, C., and Mentzer, F. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pp.\ 292--309. Springer, 2025

    2025

  43. [51]

    Improving virtual try-on with garment-focused diffusion models

    Wan, S., Li, Y., Chen, J., Pan, Y., Yao, T., Cao, Y., and Mei, T. Improving virtual try-on with garment-focused diffusion models. In European Conference on Computer Vision, pp.\ 184--199. Springer, 2024

    2024

  44. [52]

    Omnitokenizer: A joint image-video tokenizer for visual generation

    Wang, J., Jiang, Y., Yuan, Z., Peng, B., Wu, Z., and Jiang, Y.-G. Omnitokenizer: A joint image-video tokenizer for visual generation. Advances in Neural Information Processing Systems, 37: 0 28281--28295, 2024 a

    2024

  45. [53]

    Loong: Generating minute-level long videos with autoregressive language models

    Wang, Y., Xiong, T., Zhou, D., Lin, Z., Zhao, Y., Kang, B., Feng, J., and Liu, X. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024 b

    2024 arXiv

  46. [54]

    Attention is all you need

    Waswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, L., and Polosukhin, I. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  47. [55]

    Attngan: Fine-grained text to image generation with attentional generative adversarial networks

    Xu, T., Zhang, P., Huang, Q., Zhang, H., Gan, Z., Huang, X., and He, X. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1316--1324, 2018

    2018

  48. [56]

    G., Yang, M.-H., Hao, Y., Essa, I., et al

    Yu, L., Cheng, Y., Sohn, K., Lezama, J., Zhang, H., Chang, H., Hauptmann, A. G., Yang, M.-H., Hao, Y., Essa, I., et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10459--10469, 2023

    2023

  49. [57]

    B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A

    Yu, L., Lezama, J., Gundavarapu, N. B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A. G., Gong, B., Yang, M.-H., Essa, I., Ross, D. A., and Jiang, L. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth Interna...

    2024

  50. [58]

    Y., Baldridge, J., Lee, H., and Yang, Y

    Zhang, H., Koh, J. Y., Baldridge, J., Lee, H., and Yang, Y. Cross-modal contrastive learning for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 833--842, 2021

    2021

  51. [59]

    Trip: Temporal residual learning with image noise prior for image-to-video diffusion models

    Zhang, Z., Long, F., Pan, Y., Qiu, Z., Yao, T., Cao, Y., and Mei, T. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8671--8681, 2024

    2024

  52. [60]

    Towards language-free training for text-to-image generation

    Zhou, Y., Zhang, R., Chen, C., Li, C., Tensmeyer, C., Yu, T., Gu, J., Xu, J., and Sun, T. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pp.\ 17907--17917, 2022

    2022

  53. [61]

    Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis

    Zhu, M., Pan, P., Chen, W., and Yang, Y. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5802--5810, 2019

    2019

  54. [62]

    Zhu, R., Pan, Y., Li, Y., Yao, T., Sun, Z., Mei, T., and Chen, C. W. Sd-dit: Unleashing the power of self-supervised discrimination in diffusion transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8435--8445, 2024

    2024

  55. [63]

    Lumina-next : Making lumina-t2x stronger and faster with next-dit

    Zhuo, L., Du, R., Xiao, H., Li, Y., Liu, D., Huang, R., Liu, W., Zhu, X., Wang, F.-Y., Ma, Z., Luo, X., Wang, Z., Zhang, K., Zhao, L., Liu, S., Yue, X., Ouyang, W., Qiao, Y., Li, H., and Gao, P. Lumina-next : Making lumina-t2x stronger and faster with next-dit. In The Thirty-e...

    2024

  56. [64]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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