Pith. sign in

REVIEW 4 major objections 4 minor 45 references

ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal Large Language Models

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

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

Pith's one-line read Reasoning about layout relations beats raw-coordinate generation

desk verdict ReLayout is a plausible empirical improvement for LLM-based layout generation, but the paper's central claim about relation-CoT is not yet backed by a clean ablation, and the quantitative evaluation has selection and metric issues. read the letter →

arxiv 2507.05568 v1 pith:S3ZAPFH5 submitted 2025-07-08 cs.CV cs.LG

Jiaxu Tian , Xuehui Yu , Yaoxing Wang , Pan Wang , Guangqian Guo , Shan Gao This is my paper · ORCID
classification cs.CVcs.LG
keywords content-awarelayoutgenerationrelationchain-of-thoughtmulti-modallargelanguagemodelprototyperebalanceposterdiversitysaliency-aware
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 sets out to show that a multimodal large language model produces cleaner and more varied poster layouts when its training data names the spatial relationships between design elements instead of only storing their coordinates. The proposed method, ReLayout, adds relation-chain-of-thought annotations: each layout is decomposed into nested regions with a row/column direction, elements carry margin values, and salient background blocks are marked, all encoded as HTML. A second component clusters layouts into prototypes and rebalances the training distribution so rare styles are not drowned out. On the PKU and CGL poster datasets, the method lowers overlap and Fréchet distance compared with earlier LLM-based layout generators, degrades less on a hard split with nested and parallel structures, and is preferred by professional designers in user studies.

What carries the argument

The load-bearing object is the relation-CoT construction built by Algorithm 1. It projects all bounding boxes onto the x and y axes, groups overlapping projections via an IoD threshold into connected components, chooses row versus column direction by comparing group counts and variances, and recurses on each group to produce a nested region tree. Each region carries flex-direction, align-items, and margins, while salient blocks are found by an integral-image scan that scores white and black pixel density; the whole annotation is serialized as HTML in the training sequence. The second mechanism is the layout prototype rebalance sampler, which represents each layout by a weighted vector of saliency centroid features, region statistics, and element-category frequencies, clusters them with K-means into eight prototypes, and draws training samples with weight proportional to $\mathrm{cnt}^{1/\theta}$ so rare prototypes are seen more often.

What would settle it

Take a random sample of PKU and CGL posters, run Algorithm 1, and have designers reconstruct the intended region tree; if agreement on nested and parallel layouts is low, the relation-CoT labels are unreliable. A second check is to train ReLayout with the same relation annotations but shuffled margin or region assignments and show that the Ove and FD improvements disappear if the labels carry no true signal.

Watch

Extended reading notes

Core claim

The central claim is that explicit relation reasoning is the missing link between element-level coordinate prediction and layout-level design intent. ReLayout represents a layout as a recursive tree of regions, each with a flex-direction and align-items, records the margin between consecutive elements, and annotates salient background blocks that elements must not occlude. This relation-CoT annotation is used both as the model's output format, with relations predicted before coordinates, and as the basis for a prototype rebalance sampler that clusters layouts by saliency, region, and element features and resamples them to reduce long-tail bias. The paper reports that this combination beats the previous state of the art PosterLlama on overlap and Fréchet distance, transfers across datasets, and produces layouts that designers rate as usable and best.

Load-bearing premise

The automatic region-decomposition heuristic in Algorithm 1 must recover the structure a human designer intended; if it systematically mislabels layouts, the model trains on wrong relations and the reported gains would not transfer.

Editorial extensions

If this is right

  • Relation-CoT training transfers: ReLayout trained on PKU improves over PosterLlama on most metrics when tested on CGL, and vice versa, so the relation annotation is not overfit to one dataset.
  • Hard layouts benefit most: on splits containing nested regions, parallel relations, and more than four elements, ReLayout's overlap and Fréchet distance degrade less than baselines'.
  • Saliency annotations matter for content-aware quality: adding salient blocks to region-only training reduces occlusion and Fréchet distance, showing background-awareness is needed in addition to structure.
  • The rebalance sampler needs a middle setting: theta equal to 6 balances the original distribution against rare prototypes, while lower or much higher theta both hurt metrics.
  • Relation-first generation is more explainable: because the model emits region and saliency relations before bounding-box coordinates, its layout decisions can be inspected and justified in design terms.

Reading between the lines

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

  • A natural extension is to apply the same relation-CoT annotation and prototype rebalancing to other multimodal LLMs; the paper does not test this, but if the mechanism is what drives the gain, the improvement should transfer across base models.
  • The hierarchy extraction assumes layouts fit a nested row/column grammar. A stress test would be to run Algorithm 1 on free-form or diagonally placed designs; layouts that do not fit that grammar may be systematically mislabeled, which could bound the method's generality.
  • The prototype sampler is a generic recipe for de-biasing any design dataset: define interpretable layout features, cluster, and resample by inverse-frequency to the power 1/theta. One could test it on UI and document-layout datasets where long-tail arrangements are also common.
  • The relation annotations themselves could be generated automatically on unlabeled canvases using the same saliency and region detectors, supporting self-supervised pretraining and reducing the need for costly human-annotated poster datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. ReLayout proposes a content-aware layout generation method built on a multi-modal large language model (InternVL2.5-8B with LoRA fine-tuning). The method adds a 'relation-CoT' annotation layer that decomposes layouts into recursive region structures and adds saliency and margin annotations expressed in HTML-like output, followed by a 'layout prototype rebalance sampler' that clusters layouts by saliency/region/element features and reweights training samples. Experiments on the PKU and CGL datasets compare ReLayout with PosterLlama, RALF, LayoutPrompter, and a fine-tuned InternVL baseline, reporting automatic metrics, a designer user study, cross-dataset generalization, and ablations of the proposed modules.

Significance. If the central claim were fully supported, ReLayout would be a useful contribution to MLLM-based layout generation: it offers a concrete way to inject hierarchical relation structure into sequence-formatted layout outputs, and it releases two relation-annotated datasets that the community could reuse. The user study is a genuine strength, with ReLayout preferred by professional designers (Pbest 66.3% vs 12.0% for the strongest baseline) and rated usable far more often (Puse 91.0% vs 71.3%). The cross-dataset results in Table 4 also point to improved robustness. However, the paper's own ablation does not currently isolate the proposed relation-CoT mechanism, and several evaluation choices make the quantitative headline weaker than it appears. The contribution is therefore plausible but not yet convincingly demonstrated.

major comments (4)
  1. [Ablation Study and Analysis, Table 5] The only clean ablation of the relation-CoT contribution is negative: adding region annotations alone (V1) is worse than no annotations (V0) on every reported hard-split metric, with Ove rising from 0.0153 to 0.0379, FD from 8.796 to 12.229, Rea from 0.1746 to 0.1967, and Occ from 0.0821 to 0.1188. The paper itself acknowledges that the structural benefit 'cannot be quantified' because no structure-related metric is used. Since there is no V0+Resample condition, the large gains of the full method V3 (Ove 0.0097, FD 4.9403) cannot be attributed to relation reasoning rather than to the rebalance sampler or to changes in the output format. Please add (i) V0 + Resample, (ii) a saliency-only or equivalent condition that does not include region annotations, and (iii) at least one structure-aware metric (e.g., alignment consistency, nesting/parallelism validity) that can directly quantify the claimed structural benefit.
  2. [Datasets and Hyperparameter Analysis, Tables 1 and 6] The hard split is constructed using conditions that are derived from the paper's own relation ontology: (1) one region nested within another, (2) a parallel relationship, and (3) more than four elements. These are exactly the structures that relation-CoT teaches the model to emit, so the hard split is a favorable test rather than an independent stress test. Moreover, the rebalance exponent theta is selected on the hard split in Table 6, and the same hard split is then used for the headline results in Table 1. Please select hyperparameters on the regular validation split and report hard-split numbers only as a diagnostic, or use a nested selection procedure that avoids fitting theta to the reported test set.
  3. [Evaluation Metrics] The overlap metric is changed from the average IoU used in previous work (RALF, PosterLlama, and earlier content-aware methods) to the maximum IoU. A method that is explicitly designed to avoid any overlapping pair will look better under max-IoU even if its average overlap is worse. Please report both average IoU and maximum IoU, or provide empirical evidence that the two metrics do not re-rank the compared methods. Without this, the large improvement on the Ove metric in Tables 1 and 4 may partly reflect a metric choice rather than a genuine layout-quality improvement.
  4. [Layout Relation-CoT Construction, Saliency paragraph] The saliency detection algorithm is load-bearing for the content-aware claims, as it produces the salient blocks used both in training and in the Occlusion metric. In the main text it is described only as an iterative algorithm that evaluates white/black pixel density, with details deferred to the supplementary material. Please provide the full algorithm in the main text or supplement with a validation of detected salient regions against human-annotated salient objects or a standard saliency benchmark, since errors here propagate directly into the Occ metric.
minor comments (4)
  1. [Layout Relation-CoT Construction, Saliency paragraph] The sentence 'Moreover, Section also explains that adding salient blocks is crucial for the model to understand the background' contains a dangling cross-reference; please supply the correct section or figure number.
  2. [Figure 1] The header 'PosterLlamaReLayout (Ours)' reads as a single unspaced token; please fix the layout so the two method labels are clearly separated.
  3. [Table 1, CGL dataset] Please clarify why RALF achieves FD=0.9277 on the CGL test split, far below all other methods including ReLayout; if this low value is related to retrieval overlap with the training distribution, state that explicitly in the discussion.
  4. [Datasets] The redefinition of the Banner category as elements whose text/underlay IoU or IoD exceeds 0.95 changes the evaluation protocol relative to prior work; please report sensitivity to this threshold or justify the choice with examples.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the hard-split headline numbers are produced by selecting the rebalance hyperparameter on the same hard split, so the quantitative 'hard split' advantage is partly a fitted result rather than an independent prediction.

  1. fitted input called prediction [Experiments / Hyperparameter Analysis (Table 6) and Main Results (Table 1, hard split)]
    "We analyze the hyperparameter θ on the hard split of PKU dataset, and the results are shown in Table 6. It can be observed that when θ = 6, most of the metrics achieve their optimal values."

    The final hard-split row for ReLayout in Table 1 (PKU: Ove 0.0109, FD 3.4615, Rea 0.1727, Occ 0.0637) is exactly the θ=6 row of Table 6. Because θ was selected by comparing metrics on the hard split itself, and the hard split is drawn from the test and validation sets, the hard-split numbers in Table 1 are not out-of-sample predictions: they are post-selection values on the evaluation split. The reported robustness gain on the hard split is therefore statistically forced by choosing the sampler weight to optimize that exact split, rather than being an independent demonstration of relation-CoT's benefit. This is a fitted hyperparameter presented as an achieved result.

full rationale

Aside from the θ selection issue, I find no circular derivation. The relation-CoT annotations are a deterministic recoding of ground-truth boxes, and the MLLM is trained by standard supervised fine-tuning; no uniqueness theorem, ansatz-by-citation, or self-citation chain carries the argument. The IoD citation (Yu et al. 2020) is a minor non-load-bearing self-citation. The hard split is defined using the paper's own relation ontology ('one region is nested within another, a parallel relationship'), which makes that benchmark less independent, but this is a test-construction concern rather than an Equation-X=Equation-Y reduction. The ablation evidence is also not circular yet is damaging to the central attribution: Table 5 shows V1 (region annotations only) is worse than V0 on every hard-split metric, with the paper conceding 'V1 demonstrates relatively poor overall metrics... Since no structure-related metrics, this effect cannot be quantified.' Thus the claim that relation-CoT is the active ingredient is empirically under-supported, but under-support is a correctness risk, not circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central method rests on assumptions about how layouts decompose, how saliency is detected, and how prototypes cluster, plus several hand-set constants. These are not derived from external principles and are the most fragile part of the pipeline.

free parameters (5)
  • theta (rebalance exponent) = 6 (selected on hard split)
    Controls how uniform cluster sampling weights are; theta=6 chosen from values 3, 6, 10, 100 on the PKU hard split (Table 6).
  • K (number of layout prototype clusters) = 8
    Set to 8 'to maintain reasonable group sizes'; no principled selection or sensitivity analysis is given.
  • alpha, beta, gamma (feature concatenation weights) = not stated
    Equation 1 defines aggregated features as a weighted concatenation, but no values or fitting procedure are given.
  • phi (overlap threshold in GroupByOverlap) = not stated
    Algorithm 1 requires an overlap threshold for IoD grouping; no value or sensitivity analysis is given.
  • Banner category IoU/IoD threshold = 0.95
    In Datasets, text and underlay are treated as a single Banner element when IoU or IoD exceeds 0.95; this hand-chosen threshold affects the labels and all downstream metrics.
assumptions (3)
  • domain assumption Bounding boxes in a layout can be faithfully represented as a nested hierarchy of single-direction (row/column) regions with margins.
    The entire relation-CoT annotation scheme (Algorithm 1 and Figure 4) assumes layouts decompose into such regions; complex designs that mix directions or use overlapping free placement will be mislabeled.
  • domain assumption Saliency detection based on white/black pixel density approximates human visual attention.
    The saliency block construction in Layout Relation-CoT uses integral images and pixel density; no validation against human eye-tracking or saliency benchmarks is provided.
  • domain assumption K-means clustering in the three-part feature space identifies meaningful layout prototypes.
    The prototype rebalance sampler assumes that K-means over saliency centers, region statistics, and category histograms partitions designs into styles that are worth balancing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal Large Language Models." pith.science (2026). https://pith.science/paper/S3ZAPFH5

@misc{pith2026250705568,
  author       = {Pith},
  title        = {Pith review of: ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S3ZAPFH5}},
  note         = {Machine review of arXiv:2507.05568}
}
read the original abstract

Content-aware layout aims to arrange design elements appropriately on a given canvas to convey information effectively. Recently, the trend for this task has been to leverage large language models (LLMs) to generate layouts automatically, achieving remarkable performance. However, existing LLM-based methods fail to adequately interpret spatial relationships among visual themes and design elements, leading to structural and diverse problems in layout generation. To address this issue, we introduce ReLayout, a novel method that leverages relation-CoT to generate more reasonable and aesthetically coherent layouts by fundamentally originating from design concepts. Specifically, we enhance layout annotations by introducing explicit relation definitions, such as region, salient, and margin between elements, with the goal of decomposing the layout into smaller, structured, and recursive layouts, thereby enabling the generation of more structured layouts. Furthermore, based on these defined relationships, we introduce a layout prototype rebalance sampler, which defines layout prototype features across three dimensions and quantifies distinct layout styles. This sampler addresses uniformity issues in generation that arise from data bias in the prototype distribution balance process. Extensive experimental results verify that ReLayout outperforms baselines and can generate structural and diverse layouts that are more aligned with human aesthetics and more explainable.

Figures

Figures reproduced from arXiv: 2507.05568 by the authors.

Figure 1
Figure 1. Comparison with current SOTA PosterLlama (Seol, Kim, and Yoo 2024). (a) PosterLlama exhibits structural issues: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of ReLayout. We adopt the layout [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. (a) is ReLayout training process and its output distinction from previous methods. The bottom part is two key com [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples of hierarchical decomposition of com [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on the PKU and CGL [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 26 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    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. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    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. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  5. [5]

    Cai, Z.; Cao, M.; Chen, H.; Chen, K.; Chen, K.; Chen, X.; Chen, X.; Chen, Z.; Chen, Z.; Chu, P.; et al. 2024. Internlm2 technical report. arXiv preprint arXiv:2403.17297

    arXiv 2024

  6. [6]

    Chai, S.; Zhuang, L.; and Yan, F. 2023. Layoutdm: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 18349--18358

    work page 2023

  7. [7]

    Chen, Z.; Wang, W.; Cao, Y.; Liu, Y.; Gao, Z.; Cui, E.; Zhu, J.; Ye, S.; Tian, H.; Liu, Z.; et al. 2024 a . Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling. arXiv preprint arXiv:2412.05271

    arXiv 2024

  8. [8]

    Chen, Z.; Wu, J.; Wang, W.; Su, W.; Chen, G.; Xing, S.; Zhong, M.; Zhang, Q.; Zhu, X.; Lu, L.; et al. 2024 b . Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24185--24198

    2024

Show all 45 references
  1. [9]

    Deka, B.; Huang, Z.; Franzen, C.; Hibschman, J.; Afergan, D.; Li, Y.; Nichols, J.; and Kumar, R. 2017. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual ACM symposium on user interface software and technology, 845--854

    2017

  2. [10]

    Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2020. Generative adversarial networks. Communications of the ACM, 63(11): 139--144

    2020

  3. [11]

    P.; Yuan, D.; Zhao, F.; Chintagunta, J.; Naik, G.; and Ward, J

    Goyal, S.; Rastogi, E.; Rajagopal, S. P.; Yuan, D.; Zhao, F.; Chintagunta, J.; Naik, G.; and Ward, J. 2024. Healai: A healthcare llm for effective medical documentation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, 1167--1168

    2024

  4. [12]

    Guo, S.; Jin, Z.; Sun, F.; Li, J.; Li, Z.; Shi, Y.; and Cao, N. 2021. Vinci: an intelligent graphic design system for generating advertising posters. In Proceedings of the 2021 CHI conference on human factors in computing systems, 1--17

    2021

  5. [13]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

    2020

  6. [14]

    Horita, D.; Inoue, N.; Kikuchi, K.; Yamaguchi, K.; and Aizawa, K. 2024. Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 67--76

    2024

  7. [15]

    Y.; He, X.; Peng, Y.; Kong, H.; and Zhang, Q

    Hsu, H. Y.; He, X.; Peng, Y.; Kong, H.; and Zhang, Q. 2023. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6018--6026

    2023

  8. [16]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

    2022

  9. [17]

    Inoue, N.; Kikuchi, K.; Simo-Serra, E.; Otani, M.; and Yamaguchi, K. 2023. Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10167--10176

    2023

  10. [18]

    Jiang, Z.; Sun, S.; Zhu, J.; Lou, J.-G.; and Zhang, D. 2022. Coarse-to-fine generative modeling for graphic layouts. In Proceedings of the AAAI conference on artificial intelligence, volume 36, 1096--1103

    2022

  11. [19]

    A.; Durand, T.; He, J.; Sigal, L.; and Mori, G

    Jyothi, A. A.; Durand, T.; He, J.; Sigal, L.; and Mori, G. 2019. Layoutvae: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9895--9904

    2019

  12. [20]

    Kikuchi, K.; Simo-Serra, E.; Otani, M.; and Yamaguchi, K. 2021. Constrained graphic layout generation via latent optimization. In Proceedings of the 29th ACM International Conference on Multimedia, 88--96

    2021

  13. [21]

    Kingma, D. P. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  14. [22]

    Li, B.; Zhang, Y.; Guo, D.; Zhang, R.; Li, F.; Zhang, H.; Zhang, K.; Zhang, P.; Li, Y.; Liu, Z.; et al. 2024. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326

    2024 arXiv

  15. [23]

    Li, F.; Liu, A.; Feng, W.; Zhu, H.; Li, Y.; Zhang, Z.; Lv, J.; Zhu, X.; Shen, J.; Lin, Z.; et al. 2023 a . Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management,...

    2023

  16. [24]

    Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023 b . Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, 19730--19742. PMLR

    2023

  17. [25]

    Li, J.; Yang, J.; Hertzmann, A.; Zhang, J.; and Xu, T. 2019 a . LayoutGAN: Generating Graphic Layouts with Wireframe Discriminators. In International Conference on Learning Representations

    2019

  18. [26]

    Li, J.; Yang, J.; Hertzmann, A.; Zhang, J.; and Xu, T. 2019 b . Layoutgan: Generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767

    2019 arXiv

  19. [27]

    Lin, J.; Guo, J.; Sun, S.; Yang, Z.; Lou, J.-G.; and Zhang, D. 2024. Layoutprompter: Awaken the design ability of large language models. Advances in Neural Information Processing Systems, 36

    2024

  20. [28]

    Lin, J.; Zhou, M.; Ma, Y.; Gao, Y.; Fei, C.; Chen, Y.; Yu, Z.; and Ge, T. 2023. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, 1250--1260

    2023

  21. [29]

    Peng, R.; Liu, K.; Yang, P.; Yuan, Z.; and Li, S. 2023. Embedding-based retrieval with llm for effective agriculture information extracting from unstructured data. arXiv preprint arXiv:2308.03107

    2023 arXiv

  22. [30]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  23. [31]

    Raneburger, D.; Popp, R.; and Vanderdonckt, J. 2012. An automated layout approach for model-driven WIMP-UI generation. In Proceedings of the 4th ACM SIGCHI symposium on Engineering interactive computing systems, 91--100

    2012

  24. [32]

    Seol, J.; Kim, S.; and Yoo, J. 2024. PosterLlama: Bridging Design Ability of Language Model to Content-Aware Layout Generation. In European Conference on Computer Vision, 451--468. Springer

    2024

  25. [33]

    Tabata, S.; Yoshihara, H.; Maeda, H.; and Yokoyama, K. 2019. Automatic layout generation for graphical design magazines. In ACM SIGGRAPH 2019 Posters, 1--2

    2019

  26. [34]

    ???? LayoutNUWA: Revealing the Hidden Layout Expertise of Large Language Models

    Tang, Z.; Wu, C.; Li, J.; and Duan, N. ???? LayoutNUWA: Revealing the Hidden Layout Expertise of Large Language Models. In The Twelfth International Conference on Learning Representations

  27. [35]

    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. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2023 arXiv

  28. [36]

    Tzachor, A.; Devare, M.; Richards, C.; Pypers, P.; Ghosh, A.; Koo, J.; Johal, S.; and King, B. 2023. Large language models and agricultural extension services. Nature food, 4(11): 941--948

    2023

  29. [37]

    Vaswani, A. 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  30. [38]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  31. [39]

    Yang, X.; Mei, T.; Xu, Y.-Q.; Rui, Y.; and Li, S. 2016. Automatic generation of visual-textual presentation layout. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 12(2): 1--22

    2016

  32. [40]

    G.; and Wang, D

    Yang, Z.; Xu, X.; Yao, B.; Rogers, E.; Zhang, S.; Intille, S.; Shara, N.; Gao, G. G.; and Wang, D. 2024. Talk2care: An llm-based voice assistant for communication between healthcare providers and older adults. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquit...

    2024

  33. [41]

    Yu, X.; Gong, Y.; Jiang, N.; Ye, Q.; and Han, Z. 2020. Scale match for tiny person detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, 1257--1265

    2020

  34. [42]

    Zhang, J.; Guo, J.; Sun, S.; Lou, J.-G.; and Zhang, D. 2023. Layoutdiffusion: Improving graphic layout generation by discrete diffusion probabilistic models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 7226--7236

    2023

  35. [43]

    Zheng, X.; Qiao, X.; Cao, Y.; and Lau, R. W. 2019. Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG), 38(4): 1--15

    2019

  36. [44]

    Zhong, X.; Tang, J.; and Yepes, A. J. 2019. Publaynet: largest dataset ever for document layout analysis. In 2019 International conference on document analysis and recognition (ICDAR), 1015--1022. IEEE

    2019

  37. [45]

    Zhou, M.; Xu, C.; Ma, Y.; Ge, T.; Jiang, Y.; and Xu, W. 2022. Composition-aware graphic layout GAN for visual-textual presentation designs. In IJCAI

    2022

Pith tools

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