Pith. sign in

REVIEW 3 major objections 5 minor 50 references

SynVAR: Synergizing Spatial and Semantic Alignment in Visual Autoregressive Model

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SynVAR claims it can suppress cross-scale error propagation in visual autoregressive models with three training-free mechanisms.

desk verdict Genuinely VAR-specific failure analysis and a plausible three-part plug-in, but the headline gains are confounded by the extra layout input and need a clean control before the causal claim holds. read the letter →

arxiv 2608.07948 v1 pith:23WZER32 submitted 2026-08-08 cs.CV

Zhennan Chen , Tianxing Shi , Pengcheng Xu , Kepan Nan , Qian Wang , Zili Yi , Jian Yang , Ying Tai This is my paper · ORCID
classification cs.CV
keywords visualautoregressivemodeltraining-freeenhancementcomplexscenegenerationcross-scaleerrorpropagationspatial-semanticalignmenttext-to-imagehigh-frequencycompensation
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

SynVAR is a training-free enhancement framework for visual autoregressive (VAR) text-to-image models, which generate images by predicting progressively finer scale maps rather than one token at a time. The paper's central claim is that VAR's main failure in complex scenes, wrong object positions, mixed attributes, and lost details, comes from cross-scale error propagation: an error at a low-resolution scale is inherited and amplified by every later scale. It argues that diffusion-style enhancement methods cannot fix this because diffusion latents condition only on the previous step, so they are structurally mismatched with VAR's conditioning chain. SynVAR attacks the problem with three interventions: global guidance that places per-concept spatial regions into early cross-attention, receptive-field constraints that suppress cross-region attention with a Gaussian mask, and high-frequency compensation that sharpens details via a Fourier-domain filter. If the claim is right, any VAR-based text-to-image model can be improved for complex prompts without retraining; the paper reports average overall-score gains of 19.6% on Geneval and 7.9% on T2I-CompBench when SynVAR is added to Infinity and Switti.

What carries the argument

The load-bearing object is VAR's next-scale conditioning chain, $p(r_0,r_1,\dots,r_S)=\prod_s p(r_s|r_0,\dots,r_{s-1})$, in which feature maps accumulate as $f_s=f_{s-1}+\phi(r_{s-1})$. This chain is why a small early error becomes a large later error. SynVAR's machinery is a set of three interventions at the transformer level: (1) global guidance modifies early cross-attention by computing per-concept attention maps $Q_i^\top K_i$ from individual text embeddings, cropping each map to its assigned region and concatenating the crops back into the token sequence; (2) a receptive-field constraint adds a Gaussian bias $\exp(-\|p_q-p_k\|^2/(2\sigma^2))$ masked by region membership to early self-attention logits; (3) high-frequency compensation multiplies each feature map's 2D Fourier spectrum by $1+\delta\,D(u,v)$ before the inverse transform. The first two are applied at step 2, where the model's spatial and semantic choices are fixed, and the third runs at every step.

What would settle it

Run the base VAR model (Infinity or Switti) with the same per-concept bounding boxes and sub-prompts supplied as an external layout map or extra cross-attention input, but with SynVAR's Gaussian attention mask and Fourier filter disabled. If this baseline matches or beats SynVAR's Geneval Position and Attribute_Binding scores, the layout prior, not the proposed mechanisms, is causing the improvement; if it falls well below SynVAR, the internal mechanisms are doing the work.

Watch

Extended reading notes

Core claim

The paper claims that the bottleneck for VAR models in complex scenes is not capacity but cross-scale error propagation, and that a training-free intervention can suppress it. In the VAR paradigm each scale token map $r_s$ is sampled from $p(r_s|r_0,\dots,r_{s-1})$, so errors made when the image is a few pixels wide are inherited by every later scale. Experiments that swap an object's position or its semantic token show the model locks in those choices very early; attention focus also stabilizes by low resolution. SynVAR therefore intervenes at the early decision window: global guidance injects spatial boxes into cross-attention, a Gaussian region mask restrains early self-attention, and high-frequency compensation runs throughout to restore detail. The authors report that plugging this into Infinity and Switti improves overall scores by averages of 19.6% on Geneval and 7.9% on T2I-CompBench, with the largest single gain +37.0 on the Geneval Position metric for Infinity, while transplanting diffusion enhancement methods (DenseDiffusion, RAG-Diffusion) usually hurts the base models.

Load-bearing premise

Every SynVAR run in the main comparison is given concept-specific spatial boxes and sub-prompts that the vanilla baselines are not given, and no baseline gets the same layout information without SynVAR, so the reported gains could come from the injected layout prior rather than from the framework's three mechanisms.

Editorial extensions

If this is right

  • Plugging SynVAR into existing VAR text-to-image models such as Infinity and Switti improves generation quality without retraining, with average overall-score gains of 19.6% on Geneval and 7.9% on T2I-CompBench.
  • Existing diffusion enhancement methods are not directly transferable to VAR; in the paper's experiments DenseDiffusion and RAG-Diffusion generally degrade Infinity and Switti.
  • SynVAR's correction works at the early decision window: applying global guidance and receptive-field constraints at step 2 gives the best results, while applying them across many steps hurts performance.
  • High-frequency compensation adds most of the runtime overhead (2.60 to 3.10 seconds per image) but can be combined with FastVAR, which brings runtime down to 2.27 seconds per image while retaining gains.
  • The framework is not inherently dependent on a multimodal language model: automated layout planning via MLLM is used mainly for large-scale evaluation, and manual region and concept division gives the largest gains, with different MLLMs all yielding substantial improvements.

Reading between the lines

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

  • Beyond the paper, the missing control is a vanilla base model that receives the same per-concept boxes and sub-prompts without SynVAR's attention mask and frequency filter; such an experiment would separate the layout prior's contribution from the three-component mechanism.
  • Beyond the paper, the same three interventions could be tried on any autoregressive model with scale-wise conditioning, including video or multimodal generation models, since the structural argument only depends on the chain $p(r_s\mid r_{<s})$.
  • Beyond the paper, SynVAR's early-fixation diagnostic suggests a practical tuning rule: record the scale at which attention maps stop changing and set the intervention window there, instead of fixing step 2 for all prompts.
  • Beyond the paper, the fixed hyperparameters ($\sigma=0.5$, $\delta=0.01$, step 2) are per-dataset choices; a per-prompt selection mechanism would likely increase the average gain, a direction the authors list as future work.
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. SynVAR proposes a training-free enhancement framework for visual autoregressive (VAR) text-to-image models, targeting complex scenes with multiple objects/attributes. The method combines three components: (1) global guidance that decomposes the prompt into per-concept embeddings and injects them into cropped spatial regions during early generation steps; (2) a receptive field constraint that applies a Gaussian mask to self-attention in early steps to reduce cross-region semantic coupling; and (3) high-frequency compensation that boosts high-frequency components of feature maps via FFT filtering. Experiments on Geneval and T2I-CompBench with Infinity and Switti report consistent improvements, with ablations over each component, hyperparameters, early-step selection, and different MLLM planners.

Significance. If the central claim is validated, SynVAR would be a useful, low-cost plug-in for VAR-based text-to-image models, and the analysis of early fixation and cross-scale error propagation is a valuable observation for the community. The paper is well structured and includes extensive ablations, user studies, and practical extensions (overlapping regions, multiple instances, non-rectangular masks). However, the evaluation as presented does not isolate the contribution of SynVAR's internal mechanisms from the injected layout prior, so the headline causal claim is not yet established. The strength of the claimed gains—especially the +37.0 on Geneval Position—depends on resolving this confounding.

major comments (3)
  1. [3.3, Algorithm 1, Table 1] The central claim that SynVAR's three mechanisms drive the improvement is not established because the evaluation is confounded by the layout prior. In every SynVAR run, the model receives concept-specific spatial regions {h_i, w_i} and per-concept text embeddings {y_i} (Sec. 3.3, Algorithm 1), whereas the vanilla Infinity/Switti baselines receive only the raw prompt. The ablation 'w/o Global Guidance' removes the layout together with the mechanism, and 'w/o Receptive Field Constraint' retains the layout while removing only the Gaussian mask; no condition feeds the same layout to a vanilla model while withholding SynVAR's receptive-field or high-frequency operations. Consequently, the reported gains (e.g., +37.0 on Geneval Position, +21.87 on T2I-CompBench Spatial) could be caused by the injected spatial prior rather than the framework's internal mechanisms. The authors should add a control baseline that receives the same MLLM or manual layout regions and per-concept embeddings but does not apply the Gaussian mask or high-frequency compensation; if that control attains similar scores, the causal attribution to the three components would be unsupported.
  2. [8.1, Table 7] The robustness claim in Sec. 8.1 that 'the performance improvement mainly depends on our method, not the MLLM itself' does not address the confound. Table 7 compares different MLLM planners under SynVAR, but every row includes both the layout prior and the SynVAR mechanisms; there is no baseline that uses the same MLLM-produced layout without SynVAR. Therefore the table cannot distinguish the MLLM's contribution to the gain from SynVAR's contribution. A layout-only baseline is needed here as well.
  3. [4.2, Table 1, Abstract] The improvement percentages reported in the abstract (19.6% and 7.9% average improvements) do not match the numbers in Table 1 and Sec. 4.2. The Geneval overall score improves from 55.48 to 70.79 for Infinity (+27.6% relative) and from 41.09 to 45.27 for Switti (+10.2%), giving an average of about 18.9%, not 19.6%; for T2I-CompBench the overall improvements are +11.5% and +5.7%, averaging about 8.6%, not 7.9%. The authors should state the exact computation basis (relative vs absolute, average over models) and ensure consistency between the abstract, introduction, and results section.
minor comments (5)
  1. [Table 1] There is a typo in the table: 'RAG-Diffusoin' should be 'RAG-Diffusion'.
  2. [3.5, Algorithm 1] Section 3.5 states that high-frequency compensation is applied to the 'input of each transformer layer', but Algorithm 1 applies the FFT filter once per scale step, not per layer. Please clarify the exact application point.
  3. [3.3, Algorithm 1] In Algorithm 1, line 13, the crop operation should explicitly include the concatenation step from Eq. (5); otherwise the pseudocode omits how the cropped per-concept token maps are merged into r_cat.
  4. [6] The second sentence in the Limitation section is grammatically incomplete: 'for the best performance, the optimal early step of global guidance and receptive field constraints through experiments and empirical settings' should be rephrased.
  5. [4.1] The phrase 'Notably, our framework is MLLM-free' is confusing because the large-scale evaluation uses MLLM-generated layouts; consider clarifying that this refers to inference after the layout is provided, not to the evaluation pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SynVAR's performance claims are empirical benchmark measurements, not derived from the framework's own equations; the layout-prior confound is an evaluation-validity concern, not a definitional circularity.

full rationale

SynVAR is an empirical generation-time intervention; its components (global guidance, receptive-field masking, high-frequency compensation) are defined by explicit formulas (Eqs. 3-12) that do not reference the benchmark scores. The reported Geneval/T2I-CompBench improvements are measured outputs of generated images, so no prediction is equivalent to an input by construction. The only self-referential elements are citations to the authors' prior region-aware diffusion work [9] and TextCrafter [13] as contrastive baselines; these are not load-bearing for the method's derivation. The hyperparameters sigma=0.5, delta=0.01, and step=2 are selected from ablations on Geneval, and the main SynVAR pipeline always receives MLLM-produced spatial regions/concepts that vanilla baselines do not; these are legitimate threats to causal attribution and generalization, and the absence of a controlled baseline with identical layout but without SynVAR's masking/compensation weakens the internal-validity claim. However, neither issue makes an equation or fitted parameter reduce to the target result by construction, so under the paper's own evidence chain no circular step is exhibited.

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

The method relies on three fitted hyperparameters (sigma, delta, early step) and on an external layout source (MLLM or manual). The axioms are standard engineering assumptions about the model's behavior and the correctness of the intervention. No new theoretical entities are introduced.

free parameters (4)
  • sigma (receptive field decay) = 0.5
    Tuned via ablation on Geneval (Table 3); controls spatial sensitivity of the Gaussian attention mask.
  • delta (high-frequency enhancing ratio) = 0.01
    Tuned via ablation on Geneval (Table 4); controls the strength of the Fourier high-pass filter.
  • early-step index for guidance and constraint = 2
    Chosen based on ablations (Table 5) showing best performance at step 2.
  • global regions and concept division = MLLM-generated or manual
    External layout input that is essential to the global guidance component; in the main evaluation these come from an MLLM, not from SynVAR itself.
assumptions (4)
  • domain assumption Early generation steps in VAR have outsized influence on the final output, so corrections at step 2 can prevent error propagation.
    Based on the authors' experiments (Figures 4-5), used to justify applying global guidance and receptive field constraints early.
  • domain assumption The provided spatial regions and per-concept prompts accurately capture the semantic content of the input prompt.
    Required for the global guidance cross-attention injection to be beneficial; in practice it depends on MLLM quality or manual effort.
  • domain assumption Modifying self-attention masks and feature frequencies at early steps does not significantly disrupt the model's learned feature distribution.
    The method assumes the interventions are compatible with the pretrained model's distribution.
  • standard math Standard transformer and Fourier transform mathematics are correct.
    Unproved background used in Eqs. 3-12.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynVAR: Synergizing Spatial and Semantic Alignment in Visual Autoregressive Model." pith.science (2026). https://pith.science/paper/23WZER32

@misc{pith2026260807948,
  author       = {Pith},
  title        = {Pith review of: SynVAR: Synergizing Spatial and Semantic Alignment in Visual Autoregressive Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23WZER32}},
  note         = {Machine review of arXiv:2608.07948}
}
read the original abstract

VAR has gained widespread popularity due to its next-scale prediction paradigm. However, it faces substantial performance bottlenecks when handling complex scenes with multiple objects and attributes. Existing diffusion-based enhancement methods fail to adequately address the unique challenge of cross-scale error propagation and accumulation in VAR. To this end, we propose SynVAR, the first training-free enhancement framework specifically tailored for the VAR paradigm, which introduces a spatial-semantic collaborative control strategy to effectively suppress propagation error and improve generation quality. SynVAR comprises three key components: (1) Global guidance to ensure reasonable spatial structure in the early stages, (2) Receptive field constraints to mitigate early-stage semantic confusion, (3) High-frequency compensation to recover fine-grained details. Extensive quantitative and qualitative experiments demonstrate the significant improvements in the ability of SynVAR to enhance the VAR's capability for complex scene modeling.

Figures

Figures reproduced from arXiv: 2608.07948 by the authors.

Figure 1
Figure 1. (a)(b)(c): In complex scenes, existing VAR models are prone to spatial error, semantic confusion, and detail degradation. (d) Due to different formulations and mod￾eling of the image, the solutions in existing diffusion models cannot play a positive role, while our method can significantly enhance the generation performance. Abstract. VAR has gained widespread popularity due to its next-scale prediction paradigm. Ho… view at source ↗
Figure 2
Figure 2. Graphical models for diffusion and VAR inference models. Compared with the diffusion model, VAR is more prone to error accumulation and propagation. In our investigation, simply extending diffusion-based enhancement meth￾ods such as iterative refinement [2, 9] and local attention [21, 38] to VAR faces a significant degradation, as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overall Framework of SynVAR. Global guidance and receptive field constraint are completed in the early steps of generation, while high-frequency compensation is continuously applied throughout the entire generation process. Specifically, global guid￾ance introduces spatial priors into the cross-attention layers, receptive field constraint adjusts the perceptual scope of different regions in the self-attention layers… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Early fixation of semantic and spatial information. We force the replacement of semantic or spatial information in the VAR generation process, which can only affect the result in the early stage. the model’s generation process and avoid generation errors, we apply a gl…
Figure 5
Figure 5. Figure 5: Quantitative metrics at different steps. On the Geneval benchmark, the semantic and spatial related indicators are close to convergence in the early stage. In global guidance, we break the original input prompt p into a set of individual con￾cepts {c i} N i=1, where N …
Figure 6
Figure 6. Figure 6: Attention scores between different targets. the level of dependency among them remains con￾sistently high, particularly in adjacent regions. The VAR model enhances the receptive field range effectively by using two-dimensional se￾quences, which allows for more effectiv…
Figure 7
Figure 7. Figure 7: High-frequency ratio. As generation pro￾gresses, high-frequency components increase. A small intensity coefficient is used due to its sensi￾tivity. We decompose the feature map into low-frequency and high￾frequency components to better observe the evolution of image de…
Figure 8
Figure 8. Figure 8: Qualitative comparison of different methods for adapting VAR-based text-to￾image models. for usage (Notably, our framework is MLLM-free. MLLMs are only employed for large-scale evaluation and user convenience, with additional experiments de￾tailed in the Appendix). All…
Figure 9
Figure 9. Figure 9: User study on aesthetics and text-image alignment. Our approach shows a significant im￾provement. User Study. To further val￾idate the effectiveness of Syn￾VAR from a human perspec￾tive, we conducted a user study. We randomly selected 20 prompts from the Geneval and T2…
Figure 10
Figure 10. Figure 10: Qualitative analysis of global guidance, receptive field constraint and high￾frequency compensation. Effectiveness of Receptive Field Constraint. The receptive field constraint helps reduce excessive interactions between regions, preserving object generation independe…
Figure 11
Figure 11. Figure 11: Qualitative analysis of decay coefficient σ and intensity control coefficient δ. Effectiveness of High-frequency Compensation. High-frequency compen￾sation enhances fine-grained details like textures and intricate structures. While its impact on overall metrics is mod…
Figure 12
Figure 12. Figure 12: Visual analysis of the attention area during the generation process. Inference Time Cost. As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Several schematic diagrams of regional division. 7.2 Automatic Global Regional and Concept Division In order to facilitate quantitative testing and user use, we use MLLM to auto￾matically segment spatial regions and generate bounding boxes and sub-prompts from a long …
Figure 14
Figure 14. Figure 14: Automatic Global Region and Concept Division. 8 More Analysis 8.1 Effectiveness of Different MLLM To facilitate large-scale evaluation and enhance user-friendliness, we employ MLLMs to automate global region and concept division [PITH_FULL_IMAGE:figures/full_fig_p020…
Figure 15
Figure 15. Figure 15: Demonstration of our model’s flexibility in practical generation scenarios. (a) Handling overlapping and fine-grained interactive bounding boxes. (b) Seamless generation of multiple instances belonging to the same concept (e.g., three distinct apples) within a single …
Figure 16
Figure 16. Figure 16: Different mask strategies in receptive field constraints. As shown in [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]
Figure 17
Figure 17. Figure 17: More qualitative results of SynVAR [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: More qualitative results of SynVAR [PITH_FULL_IMAGE:figures/full_fig_p025_18.png]
Figure 19
Figure 19. Figure 19: SynVAR generation results using different seeds show more stable generation results [PITH_FULL_IMAGE:figures/full_fig_p026_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 11 canonical work pages

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    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]

    Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation (2023)

    2023

  3. [3]

    Advances in neural information processing systems33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

    2020

  4. [4]

    arXiv preprint arXiv:2301.00704 (2023)

    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. arXiv preprint arXiv:2301.00704 (2023)

    arXiv 2023

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked genera- tive image transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11315–11325 (2022)

    2022

  6. [6]

    ACM trans- actions on Graphics (TOG)42(4), 1–10 (2023)

    Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., Cohen-Or, D.: Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM trans- actions on Graphics (TOG)42(4), 1–10 (2023)

    2023

  7. [7]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Chen, M., Laina, I., Vedaldi, A.: Training-free layout control with cross-attention guidance. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 5343–5353 (2024)

    2024

  8. [8]

    In: ECAI 2023, pp

    Chen, Z., Gao, R., Xiang, T.Z., Lin, F.: Diffusion model for camouflaged object detection. In: ECAI 2023, pp. 445–452. IOS Press (2023)

    work page 2023

Show all 50 references
  1. [9]

    arXiv preprint arXiv:2411.06558 (2024)

    Chen, Z., Li, Y., Wang, H., Chen, Z., Jiang, Z., Li, J., Wang, Q., Yang, J., Tai, Y.: Region-aware text-to-image generation via hard binding and soft refinement. arXiv preprint arXiv:2411.06558 (2024)

    2024 arXiv

  2. [10]

    arXiv preprint arXiv:2605.12013 (2026)

    Chen, Z., Zhu, J., Chen, X., Zhang, J., Chen, J., Zeng, Z., Zhang, W., Wang, C., Yang, J., Tai, Y.: L2p: Unlocking latent potential for pixel generation. arXiv preprint arXiv:2605.12013 (2026)

    2026 arXiv

  3. [11]

    arXiv preprint arXiv:2511.18822 (2025)

    Chen, Z., Zhu, J., Chen, X., Zhang, J., Hu, X., Zhao, H., Wang, C., Yang, J., Tai, Y.: Dip: Taming diffusion models in pixel space. arXiv preprint arXiv:2511.18822 (2025)

    2025

  4. [12]

    arXiv preprint arXiv:2412.14169 (2024)

    Deng, H., Pan, T., Diao, H., Luo, Z., Cui, Y., Lu, H., Shan, S., Qi, Y., Wang, X.: Autoregressive video generation without vector quantization. arXiv preprint arXiv:2412.14169 (2024)

    2024 arXiv

  5. [13]

    arXiv preprint arXiv:2503.23461 (2025)

    Du, N., Chen, Z., Chen, Z., Gao, S., Chen, X., Jiang, Z., Yang, J., Tai, Y.: Textcrafter: Accurately rendering multiple texts in complex visual scenes. arXiv preprint arXiv:2503.23461 (2025)

    2025

  6. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Esser, P., Rombach, R., 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

  7. [15]

    Advances in Neural Information Processing Systems36, 52132–52152 (2023)

    Ghosh, D., Hajishirzi, H., Schmidt, L.: Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems36, 52132–52152 (2023)

    2023

  8. [16]

    arXiv preprint arXiv:2503.23367 (2025) SynVAR 17

    Guo, H., Li, Y., Zhang, T., Wang, J., Dai, T., Xia, S.T., Benini, L.: Fastvar: Linear visual autoregressive modeling via cached token pruning. arXiv preprint arXiv:2503.23367 (2025) SynVAR 17

    2025 arXiv

  9. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Guo, X., Liu, J., Cui, M., Li, J., Yang, H., Huang, D.: Initno: Boosting text- to-image diffusion models via initial noise optimization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9380– 9389 (2024)

    2024

  10. [18]

    arXiv preprint arXiv:2412.04431 (2024)

    Han, J., Liu, J., Jiang, Y., Yan, B., Zhang, Y., Yuan, Z., Peng, B., Liu, X.: Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431 (2024)

    2024 arXiv

  11. [19]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  12. [20]

    Advances in Neural Information Processing Systems36, 78723–78747 (2023)

    Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems36, 78723–78747 (2023)

    2023

  13. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kim, Y., Lee, J., Kim, J.H., Ha, J.W., Zhu, J.Y.: Dense text-to-image genera- tion with attention modulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7701–7711 (2023)

    2023

  14. [22]

    arXiv preprint arXiv:2408.02657 (2024)

    Liu, D., Zhao, S., Zhuo, L., Lin, W., Qiao, Y., Li, H., Gao, P.: Lumina-mgpt: Illu- minate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657 (2024)

    2024 arXiv

  15. [23]

    arXiv preprint arXiv:2406.10797 (2024)

    Ma, X., Zhou, M., Liang, T., Bai, Y., Zhao, T., Chen, H., Jin, Y.: Star: Scale- wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797 (2024)

    2024 arXiv

  16. [24]

    arXiv preprint arXiv:2407.02371 (2024)

    Nan, K., Xie, R., Zhou, P., Fan, T., Yang, Z., Chen, Z., Li, X., Yang, J., Tai, Y.: Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371 (2024)

    2024 arXiv

  17. [25]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195–4205 (2023)

    2023

  18. [26]

    arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

    2022 arXiv

  19. [27]

    Advances in neural information processing systems32(2019)

    Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems32(2019)

    2019

  20. [28]

    Advances in neural information processing systems35, 36479–36494 (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text- to-image diffusion models with deep language understanding. Advances in neural information processing systems35,...

    2022

  21. [29]

    arXiv preprint arXiv:2010.02502 (2020)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  22. [30]

    arXiv preprint arXiv:2406.06525 (2024)

    Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., Yuan, Z.: Autoregres- sive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525 (2024)

    2024 arXiv

  23. [31]

    Tian, K., Jiang, Y., Yuan, Z., Peng, B., Wang, L.: Visual autoregressive modeling: Scalableimagegenerationvianext-scaleprediction.Advancesinneuralinformation processing systems37, 84839–84865 (2024)

    2024

  24. [32]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  25. [33]

    arXiv preprint arXiv:2307.09288 (2023) 18 Z

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 18 Z. Chen et al

    2023 arXiv

  26. [34]

    Advances in neural information processing systems30(2017)

    Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems30(2017)

    2017

  27. [35]

    arXiv preprint arXiv:2412.01819 (2024)

    Voronov,A.,Kuznedelev,D.,Khoroshikh,M.,Khrulkov,V.,Baranchuk,D.:Switti: Designing scale-wise transformers for text-to-image synthesis. arXiv preprint arXiv:2412.01819 (2024)

    2024 arXiv

  28. [36]

    arXiv preprint arXiv:2503.23897 (2025)

    Wang, Y., Guo, L., Li, Z., Huang, J., Wang, P., Wen, B., Wang, J.: Training- free text-guided image editing with visual autoregressive model. arXiv preprint arXiv:2503.23897 (2025)

    2025 arXiv

  29. [37]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Xie, J.,Li,Y., Huang,Y.,Liu, H.,Zhang,W., Zheng,Y.,Shou, M.Z.:Boxdiff: Text- to-image synthesis with training-free box-constrained diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7452–7461 (2023)

    2023

  30. [38]

    In: Forty- first International Conference on Machine Learning (2024)

    Yang, L., Yu, Z., Meng, C., Xu, M., Ermon, S., Cui, B.: Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In: Forty- first International Conference on Machine Learning (2024)

    2024

  31. [39]

    Authorea Preprints (2025)

    Yang, S., Chen, Z., Lin, Y., Chen, X., Cai, G., Yu, H., Wu, P., Yang, Q.: A sur- vey on vision-language models for multimodal federated learning tasks. Authorea Preprints (2025)

    2025

  32. [40]

    In: European Conference on Computer Vision

    Yang, Z., Jiang, Z., Li, X., Zhou, H., Dong, J., Zhang, H., Du, Y.: D 4-vton: Dynamic semantics disentangling for differential diffusion based virtual try-on. In: European Conference on Computer Vision. pp. 36–52. Springer (2024)

    2024

  33. [41]

    In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

    Yang, Z., Li, Y., He, S., Li, X., Xu, Y., Dong, J., Du, Y.: Omnivton: Training-free universal virtual try-on. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 16702–16711 (2025)

    2025

  34. [42]

    arXiv preprint arXiv:2606.27880 (2026)

    Yang, Z., Tai, Y., Zhan, J., Zheng, Y., Qian, J., Yang, J.: Orthotryon: Geomet- ric orthogonalization for conflict-free unified fashion generation. arXiv preprint arXiv:2606.27880 (2026)

    2026 arXiv

  35. [43]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhao, C., Cai, W., Dong, C., Hu, C.: Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8281–8291 (2024)

    2024

  36. [44]

    arXiv preprint arXiv:2602.11564 (2026)

    Zhao, C., Chen, J., Li, H., Kang, Z., Lu, S., Wei, X., Zhang, K., Yang, J., Tai, Y.: Luve: Latent-cascaded ultra-high-resolution video generation with dual frequency experts. arXiv preprint arXiv:2602.11564 (2026)

    2026 arXiv

  37. [45]

    IEEE Transactions on Geoscience and Remote Sensing (2026)

    Zhao, C., Dong, C., Cai, W., Wang, Y.: Learning a physical-aware diffusion model based on transformer for underwater image enhancement. IEEE Transactions on Geoscience and Remote Sensing (2026)

    2026

  38. [46]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    Zhao, C., Xu, Y., Chen, Z., Gu, E., Zhang, K., Liu, X., Yang, J., Tai, Y.: From zero to detail: A progressive spectral decoupling paradigm for uhd image restora- tion with new benchmark. IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    2026

  39. [47]

    arXiv preprint arXiv:2503.12885 (2025)

    Zhou, D., Li, M., Yang, Z., Yang, Y.: Dreamrenderer: Taming multi- instance attribute control in large-scale text-to-image models. arXiv preprint arXiv:2503.12885 (2025)

    2025 arXiv

  40. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhou, D., Li, Y., Ma, F., Zhang, X., Yang, Y.: Migc: Multi-instance generation con- troller for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6818–6828 (2024)

    2024

  41. [49]

    arXiv preprint arXiv:2410.12669 (2024) SynVAR 19

    Zhou, D., Xie, J., Yang, Z., Yang, Y.: 3dis: Depth-driven decoupled instance syn- thesis for text-to-image generation. arXiv preprint arXiv:2410.12669 (2024) SynVAR 19

    2024

  42. [50]

    A blonde ponytail European girl in a white shirt

    Zhuang, X., Xie, Y., Deng, Y., Liang, L., Ru, J., Yin, Y., Zou, Y.: Vargpt: Unified understanding and generation in a visual autoregressive multimodal large language model. arXiv preprint arXiv:2501.12327 (2025) 7 More Model Details 7.1 Explanation of Global Region Division In...

    2025 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.