Pith. sign in

REVIEW 5 major objections 6 minor 85 references

Why Settle for Mid: A Probabilistic Viewpoint to Spatial Relationship Alignment in Text-to-image Models

T0 review · 5 major / 6 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 The paper claims that replacing bounding-box-center comparisons with a whole-object probabilistic measure, Probability of Superiority, both evaluates and improves spatial relationship accuracy in text-to-image models.

desk verdict PSE is a credible, well-motivated evaluation metric and the core idea is worth taking seriously, but the headline PSG gains are oversold by a prompt-simplified variant and the gradient guidance lacks causal evidence. read the letter →

arxiv 2506.23418 v1 pith:VPBQQPVN submitted 2025-06-29 cs.CV

Parham Rezaei , Arash Marioriyad , Mahdieh Soleymani Baghshah , Mohammad Hossein Rohban This is my paper
classification cs.CV
keywords spatialrelationshipstext-to-imagegenerationprobabilityofsuperioritycompositionalevaluationmetricinference-timeguidancecross-attentionmapsdiffusionmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the standard way of checking spatial relationships in text-to-image models—comparing the bounding-box centers of two objects—is the wrong lens, because it ignores object shape and size and flips to a confident 0/1 verdict on scenes humans read as ambiguous. It replaces that lens with a probabilistic one: for a relation like "to the right of," ask what fraction of the pixels of one object lie to the right of the pixels of the other, using the Probability of Superiority ($\mathrm{PoS}$) of the two pixel distributions. On this idea it builds two tools: PSE, an evaluation metric that matches human judgments substantially better than center-based and embedding-based metrics, and PSG, a training-free inference-time guidance that nudges cross-attention maps (or selects initial noise) to raise the PoS reward, improving spatial accuracy on VISOR, T2I-CompBench, and HRS benchmarks without fine-tuning. If correct, the paper shows that a single probabilistic measure can serve as both judge and guide for spatial composition in generative models.

What carries the argument

The carrying object is the Probability of Superiority ($\mathrm{PoS}$) between two one-dimensional probability distributions, $\mathrm{PoS}(A,B) = P_{X\sim A, Y\sim B}(X > Y)$, i.e., the probability that a random draw from $A$ exceeds a random draw from $B$. For spatial relations, the two distributions are normalized projections of each object's pixel mask (or cross-attention map) onto the axis of the relation, and the PSE score subtracts the reverse-relation PoS and takes the positive part so ambiguous placements score near zero instead of a hard 0/1. In PSG this same quantity is differentiated with respect to the latent to push probability mass in the desired direction during early denoising steps, directly repurposing the mathematics of the metric as the objective of the guidance.

What would settle it

Run a human study on an independent set of, say, 1,000 generated images with spatial prompts and check whether PSE's Spearman correlation advantage over VISOR—about 0.2 in the paper—reproduces with fresh annotators and fresh prompts; if the advantage shrinks or vanishes, the metric's human-alignment claim is not robust. A complementary check is to apply gradient-based PSG to a model whose cross-attention maps are corrupted (e.g., spatially shuffled), and see whether spatial accuracy still improves; if it does, the effectiveness is not actually mediated by attention maps.

Watch

Extended reading notes

Core claim

The central claim is that spatial relationship accuracy in text-to-image models is better measured and better controlled by the Probability of Superiority of the two objects' full pixel distributions than by the relative positions of their bounding-box centers. Concretely, the paper defines $\mathrm{PSE}(A,B;r) = [\mathrm{PoS}_v(A,B) - \mathrm{PoS}_{-v}(A,B)]_+$, the positive part of the difference between the probability that a random pixel of $A$ lies on the requested side of a random pixel of $B$ and the probability of the opposite relation, computed on segmentation masks in 2D and on depth-quantized masks for 3D. This continuous score avoids the false confidence of center-based metrics on wide objects and on ambiguous middle placements. The same score, applied to cross-attention maps during denoising, becomes the reward for PSG: a gradient step on the latent (generative semantic nursing) or a Best-of-N selection over initial noise vectors. The paper reports that PSE raises Spearman correlation with human judgments from 0.551 (VISOR) to 0.726, and that PSG raises the VISOR unconditional score of SDXL from 65.74% to 80.02% while improving T2I-CompBench and HRS spatial scores.

Load-bearing premise

The gradient guidance assumes that moving probability mass in the cross-attention map of an object actually moves that object's final rendered location, without erasing or distorting it, and this assumption is supported mainly by benchmark numbers and qualitative examples rather than by a causal analysis.

Editorial extensions

If this is right

  • PSE can replace or complement center-based metrics in existing benchmarks such as VISOR, T2I-CompBench, and HRS, giving a graded, human-aligned score instead of a binary verdict.
  • PSG improves spatial accuracy across U-Net backbones (SD 1.4, SDXL) and a transformer backbone (PixArt-$\alpha$) with no fine-tuning, and the Best-of-N variant works on black-box models like FLUX, SANA, and SD3.
  • The OPSE bandit version identifies the best model with far fewer generated images, lowering the cost of model comparison when APIs are expensive.
  • Because PSG operates on the prompt alone, it avoids the layout bias of LLM-generated bounding-box guidance and produces more diverse object placements.

Reading between the lines

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

  • The same PoS machinery could be applied to other compositional attributes, such as attribute binding or size relations, whenever the attribute can be expressed as a distribution over pixels or attention mass.
  • The search-based PSG variant is a general inference-time-scaling wrapper: any differentiable or non-differentiable reward computed on the final image could be swapped into the Best-of-N loop, so PSG's 3D gains may be attributable as much to selection as to the PoS reward itself.
  • A testable extension is to use PSE as a reward in fine-tuning rather than only inference-time guidance; the paper's own limitation notes that training-free guidance may not suffice for prompts with many entities, so a PoS-based loss for fine-tuning would be a natural next step.
  • The metric's dependence on segmentation and depth estimators means its human-alignment claims are bracketed by detector quality; improvements in grounding models should transfer directly to PSE without changing the framework.
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

5 major / 6 minor

Summary. The paper proposes a probabilistic framework, based on Probability of Superiority (PoS), for evaluating and improving spatial relationships in text-to-image models. It introduces PSE, a metric that compares whole-object segmentation masks (and depth, for 3D relations) rather than bounding-box centers; PSG, an inference-time guidance method that uses a PoS-based reward either through gradient updates on cross-attention maps or through Best-of-N selection; and OPSE, a multi-armed-bandit version of PSE for efficient online model evaluation. The authors report that PSE correlates more strongly with human judgment than VISOR, T2I-CompBench, HRS-Bench, CLIP Score, XVLM, and Image Reward, and that PSG substantially improves spatial accuracy on the VISOR, T2I-CompBench, and HRS benchmarks without fine-tuning.

Significance. If the claims hold, the paper makes a useful contribution: a metric that accounts for object shape and extent rather than only centers is a principled improvement, and a training-free guidance method for spatial relations is practically valuable. The paper includes several strengths: the mathematical formulation is simple and correct; evaluation is checked against external benchmarks (VISOR, T2I-CompBench, HRS) and a fresh human study; robustness experiments show PSE is stable to mask corruption and detector choice; and PSG is demonstrated on multiple backbones, including a transformer-based model. The main weaknesses are experimental: missing uncertainty quantification, a potentially confounded human-evaluation setup, and a prompt-simplified variant whose evaluation protocol is ambiguous and which appears to drive the headline numbers.

major comments (5)
  1. [Section 4.2, Table 4] The PSG(2) variant replaces explicit spatial terms with "and" in the prompt, yet the reported VISOR numbers presumably still evaluate the original spatial relation. If so, the method is scored on a relation that is absent from the generation prompt, so the 32% headline improvement does not measure instruction-following; if instead the simplified prompt is used for evaluation, then the relation is absent from the test condition and the metric is not well defined. Please clarify exactly which prompt is used for scoring and either drop PSG(2) from the headline claims or report PSG(1) (which does not modify the prompt) as the primary result.
  2. [Section 4.1, Table 1] No confidence intervals, bootstrap estimates, or significance tests are reported for the correlation coefficients in Table 1, and the human study uses only three annotators with no inter-annotator agreement statistic. Without these, the claimed "approximately 0.2 improvement in Spearman correlation over VISOR" cannot be distinguished from sampling noise. Please add CIs (e.g., bootstrap) for all correlations and report annotator agreement (e.g., Fleiss' kappa or pairwise agreement).
  3. [Section 4.1] The human-evaluation set of 2,000 images includes 500 images generated by "SDXL + PSG," a model that is itself guided by the PoS objective. Including PSE-guided images in the metric-validation set makes the evaluation partially circular: the metric is being tested on images that were selected or steered to have high PSE scores. Please report the correlations separately for each source model, or restrict the main comparison to the three independent backbones, to establish that PSE's advantage is not an artifact of this confound.
  4. [Section 3.2, Eq. (6)] The gradient-based PSG updates the latent so that the cross-attention maps Ai and Aj move toward the requested PoS relation, but the paper does not establish that this translates into relocated final object masks rather than merely reshaping attention maps or eroding object identity. The qualitative attention maps in Fig. 5 and aggregate VISOR numbers in Table 4 are suggestive but not causal evidence. Please provide a per-prompt analysis linking the attention shift to the final segmentation (e.g., tracking mask IoU or center displacement), or an ablation that controls for object-erasure and overlap, to support the claim that PSG improves final spatial alignment.
  5. [Appendix B.1, Section 4.2] The hyperparameters of gradient-based PSG (number of guided steps and loss scale) are selected using the VISOR metric on a small development set and then evaluated on VISOR in Table 4. This creates a risk of benchmark overfitting. Please report sensitivity of the main results to these hyperparameters and, if possible, validate the chosen settings on a held-out set of relations not used for tuning.
minor comments (6)
  1. [Abstract] The abstract refers to "Part-of-Speech PoS-based reward function," but PoS stands for Probability of Superiority; please correct this typo.
  2. [Section 3.1, Eq. (5)] Please define the bracket notation [x]_+ explicitly as max(x, 0) when PSE is introduced.
  3. [Section 4.2] The sentence "5,000 prompts with four random seeds to where used compare PSG with re-evaluated baseline models" is garbled and should be rewritten.
  4. [Appendix A.1, Figure 7] The axes and tick labels in Figure 7 appear corrupted or missing; the plots should be regenerated with readable labels.
  5. [Appendix A.3, Table 11] The model name "Gounding-Dino" is a typo for "Grounding DINO."
  6. [Appendix A.3, Table 10] The robustness table reports mean PSE scores under corruptions but no standard deviation; adding variance would make the robustness claim more convincing.

Circularity Check

1 steps flagged · score 3.0 of 10

The central metric and generation claims are externally benchmarked, but the appendix best-of-N table reports PSE improvement that is guaranteed by the PSE-based selection objective; this is localized and not load-bearing.

  1. self definitional [Section 3.2 (Search-Based PSG) and Appendix B.2, Table 14]
    "We adopt this approach by using the PSE score as a reward function to guide the selection process. Specifically, we generate N = 32 samples per prompt and select the image that achieves the highest PSE score, as computed by the evaluation pipeline described in section 3.1. ... Table 14: PSE/(center position accuracy) of PSG inference time scaling. This table demonstrates the effectiveness of our method for different models. Increasing N, shows that by using more compute we can get consistently better alignment."

    The PSE column in Table 14 is computed with the same PSE metric that the selection procedure maximizes. For N > 1, choosing the highest-PSE sample among N independent draws makes the selected PSE increase with N by construction, so reporting this increase as evidence of better spatial alignment is a property of taking a maximum, not of improved object placement. The companion center-position accuracy column is an independent check and does improve, and the paper's headline claims rest on VISOR, T2I-CompBench, and HRS rather than on this table. The circularity is therefore real but localized.

full rationale

The paper's main evaluation claims are checked against external benchmarks and a fresh human study: PSE is correlated against human labels on 2,000 generated images, and PSG is compared with VISOR, T2I-CompBench, and HRS-Benchmark, none of which use the proposed PSE reward as their scoring function. The authors also explicitly mark Table 3 as 'not a performance validation for PSG as it uses the same measure,' which is the correct treatment of a self-referential comparison. Self-citations to Marioriyad et al. (2024) and Rezaei et al. (2025) are used for a failure taxonomy and bandit selection inspiration, respectively, and neither is load-bearing for the core derivation. The one construction-reduced piece of evidence is the appendix best-of-N table, where the PSE score is both the selection objective and the reported outcome; this is a genuine but minor circularity because the same table also reports center-position accuracy and the central claims do not depend on it. Overall, the derivation chain is substantially self-contained, so the circularity score is low.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new particles, forces, dimensions, or physical entities are introduced. The only new objects are the PSE metric and PSG procedure, which are mathematical artifacts, not invented entities in the ledger sense.

free parameters (7)
  • PSE binary threshold = 0.5
    Selected to match non-thresholded PSE on What'sUp and COCO-Spatial datasets (Appendix A.1).
  • PSG gradient scale factor = 1000 for SDXL, 25 for SD 1.4
    Grid-searched on 100 pairs from 10 non-COCO objects using the VISOR metric (Appendix B.1).
  • PSG guided denoising steps = 10 steps for SDXL, 20 for SD 1.4
    Same hyperparameter search as the scale factor (Appendix B.1).
  • Search-based PSG sample count = N = 32
    Set in Section 3.2 without a stated ablation; larger N monotonically improves alignment but costs compute (Appendix B.2).
  • UCB exploration constant = alpha = 2
    Set empirically for OPSE (Appendix A.6).
  • Distance loss threshold = c = 5
    Used for distance-aware PoS on 16x16 attention maps (Appendix B.4).
  • 3D IoU detection threshold = 0.5
    Used in the depth-focused 3D spatial assessment (Section 4.2).
assumptions (5)
  • standard math Probability of Superiority as defined by Wolfe and Hogg (1971) is a valid measure of ordinal dominance between two distributions.
    The whole framework, eq. (2), is the definition of PoS from the cited statistics literature.
  • domain assumption Cross-attention maps in diffusion T2I models are normalized and can be treated as probability distributions over spatial locations for each object.
    Stated in Section 3.2: "These normalized cross-attention maps can be interpreted as probability distributions over spatial locations".
  • domain assumption Gradient updates on the latent with respect to the PoS loss on attention maps shift the final object layout without breaking image quality.
    Relied on in eq. (6); supported only by qualitative examples in fig. 5 and benchmark results, not by a causal or robustness analysis.
  • domain assumption Grounded SAM masks and Depth Anything depth estimates are accurate enough proxies for the true object regions.
    Used throughout PSE; the paper's own Appendix A.5 lists this as a limitation.
  • domain assumption Human annotations in the user study are the ground truth for spatial correctness.
    The correlations in Table 1 use aggregate labels from three annotators as the reference; no inter-annotator agreement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Settle for Mid: A Probabilistic Viewpoint to Spatial Relationship Alignment in Text-to-image Models." pith.science (2026). https://pith.science/paper/VPBQQPVN

@misc{pith2026250623418,
  author       = {Pith},
  title        = {Pith review of: Why Settle for Mid: A Probabilistic Viewpoint to Spatial Relationship Alignment in Text-to-image Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VPBQQPVN}},
  note         = {Machine review of arXiv:2506.23418}
}
read the original abstract

Despite the ability of text-to-image models to generate high-quality, realistic, and diverse images, they face challenges in compositional generation, often struggling to accurately represent details specified in the input prompt. A prevalent issue in compositional generation is the misalignment of spatial relationships, as models often fail to faithfully generate images that reflect the spatial configurations specified between objects in the input prompts. To address this challenge, we propose a novel probabilistic framework for modeling the relative spatial positioning of objects in a scene, leveraging the concept of Probability of Superiority (PoS). Building on this insight, we make two key contributions. First, we introduce a novel evaluation metric, PoS-based Evaluation (PSE), designed to assess the alignment of 2D and 3D spatial relationships between text and image, with improved adherence to human judgment. Second, we propose PoS-based Generation (PSG), an inference-time method that improves the alignment of 2D and 3D spatial relationships in T2I models without requiring fine-tuning. PSG employs a Part-of-Speech PoS-based reward function that can be utilized in two distinct ways: (1) as a gradient-based guidance mechanism applied to the cross-attention maps during the denoising steps, or (2) as a search-based strategy that evaluates a set of initial noise vectors to select the best one. Extensive experiments demonstrate that the PSE metric exhibits stronger alignment with human judgment compared to traditional center-based metrics, providing a more nuanced and reliable measure of complex spatial relationship accuracy in text-image alignment. Furthermore, PSG significantly enhances the ability of text-to-image models to generate images with specified spatial configurations, outperforming state-of-the-art methods across multiple evaluation metrics and benchmarks.

Figures

Figures reproduced from arXiv: 2506.23418 by the authors.

Figure 1
Figure 1. Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison with the state-of-the-art SDXL model (Podell et al., 2023) and the fine [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Process of PSE metric for the prompt “A cat to the right of a bowl”. The depth detection is only applied for 3D relations. To quantify spatial relationships from a probabilistic perspective, we start with a simple case. Consider two distributions, A and B, which, with probability 1, take on single point values x and y, respectively. In this scenario, determining whether A is to the right of B reduces to evaluating t… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Inconsistency of center-based evaluation with human perception of spatial relationships for the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Results of applying gradient-based PSG to Stable Diffusion 1.4 on cross-attention maps of dog and cat for the prompt “A dog to the left of a Cat”. The distribution of attention of each object is gradually shifted to its correct relative position. Gradient-Based PSG (Ge…
Figure 6
Figure 6. Figure 6: Number of times each model is chosen by the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Thresholded version of PSE for different threshold values on the COCO-Spatial and What’sUp datasets. A.2 Visual Clarification of Ineffectiveness of Using Centers fig. 4 illustrates that relying on the center of bounding boxes, as done by other detection-based evaluator…
Figure 8
Figure 8. Figure 8: A failure case of VI￾SOR metric. VISOR incorrectly determines the chair to the left of the TV, while the chair is nei￾ther to the right nor left of the TV [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: fig. 9. Through experimentation, we determined that a configuration of (10 steps, 1000 scale factor) offered [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 9
Figure 9. Figure 9: Qualitative analysis of the effect of the scale factor of [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: The effect of the scale factor in the PSG method. As the scale increases, the generated images can become more distorted. Using a medium scale instead of a very small one can prevent incomplete images. (a) 1000 steps (b) 1900 steps [PITH_FULL_IMAGE:figures/full_fig_p…
Figure 11
Figure 11. Figure 11: The effect of optimization step counts in the [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Order bias in generated images with PSG variants applied on SD 1.4 and SDXL backbones. In all experiments, Object A refers to the first object and Object B refers to the second object. Relation Split Scores. We have presented the Object Accuracy and VISOR score for ea…
Figure 13
Figure 13. Figure 13: VISOR unconditional scores of our models and baselines split by super-category pairs. We observe [PITH_FULL_IMAGE:figures/full_fig_p030_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p031_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p031_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p032_16.png]
Figure 17
Figure 17. Figure 17: Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p033_17.png]
Figure 18
Figure 18. Figure 18: Qualitative comparison of our proposed method, [PITH_FULL_IMAGE:figures/full_fig_p033_18.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 34 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    A-star: Test-time attention segregation and retention for text-to-image synthesis

    Aishwarya Agarwal, Srikrishna Karanam, K J Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 2283--2293, October 2023

    2023

  3. [3]

    Kandinsky 3.0 technical report, 2024

    Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report, 2024. URL https://arxiv.org/abs/2312.03511

    arXiv 2024

  4. [4]

    Finite-time analysis of the multiarmed bandit problem

    Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47 0 (2/3): 0 235–256, 2002. ISSN 0885-6125. doi:10.1023/a:1013689704352. URL http://dx.doi.org/10.1023/A:1013689704352

    doi:10.1023/a:1013689704352 2002

  5. [5]

    Spatext: Spatio-textual representation for controllable image generation

    Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 18370–18380. IEEE, June 2023. doi:10.1109/cvpr52729.2023.01762. URL http://d...

    arXiv 2023

  6. [6]

    Cc3d: Layout-conditioned generation of compositional 3d scenes

    Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas Guibas, and Andrea Tagliasacchi. Cc3d: Layout-conditioned generation of compositional 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 7171--7181, 2023

    2023

  7. [7]

    Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models, 2023

    Eslam Mohamed Bakr, Pengzhan Sun, Xiaoqian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. Hrs-bench: Holistic, reliable and scalable benchmark for text-to-image models, 2023

    2023

  8. [8]

    Multidiffusion: Fusing diffusion paths for controlled image generation, 2023

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. URL https://arxiv.org/abs/2302.08113

    arXiv 2023

Show all 85 references
  1. [9]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. URL https://api....

  2. [10]

    Make it count: Text-to-image generation with an accurate number of objects

    Lital Binyamin, Yoad Tewel, Hilit Segev, Eran Hirsch, Royi Rassin, and Gal Chechik. Make it count: Text-to-image generation with an accurate number of objects. arXiv preprint arXiv:2406.10210, 2024

    2024 arXiv

  3. [11]

    Conceptual 12M : Pushing web-scale image-text pre-training to recognize long-tail visual concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M : Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021

    2021

  4. [12]

    Getting it right: Improving spatial consistency in text-to-image models

    Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. arXiv preprint arXiv:2404.01197, 2024

    2024 arXiv

  5. [13]

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42: 0 1 -- 10, 2023. URL https://api.semanticscholar.org/CorpusID:256416326

    2023

  6. [14]

    Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024 a

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024 a . URL https://arxiv.org/abs/2401.12168

    2024 arXiv

  7. [15]

    Pixart- : Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart- : Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023

    arXiv

  8. [16]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. 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 b

    2024

  9. [17]

    Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation, 2023

    Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation, 2023. URL https://arxiv.org/abs/2302.08908

    2023 arXiv

  10. [18]

    Dall·e mini, 7 2021

    Boris Dayma, Suraj Patil, Pedro Cuenca, Khalid Saifullah, Tanishq Abraham, Phuc Le Khac, Luke Melas, and Ritobrata Ghosh. Dall·e mini, 7 2021. URL https://github.com/borisdayma/dalle-mini

    2021

  11. [19]

    Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022

    Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022. URL https://arxiv.org/abs/2204.14217

    2022 arXiv

  12. [20]

    Training-free structured diffusion guidance for compositional text-to-image synthesis

    Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In The Eleventh International Conference on Learni...

    2023

  13. [21]

    Make-a-scene: Scene-based text-to-image generation with human priors, 2022

    Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors, 2022. URL https://arxiv.org/abs/2203.13131

    2022 arXiv

  14. [22]

    Geneval: An object-focused framework for evaluating text-to-image alignment, 2023

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment, 2023. URL https://arxiv.org/abs/2310.11513

    2023 arXiv

  15. [23]

    Benchmarking spatial relationships in text-to-image generation

    Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, and Yezhou Yang. Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015, 2022

    2022 arXiv

  16. [24]

    Gradient guidance for diffusion models: An optimization perspective, 2024

    Yingqing Guo, Hui Yuan, Yukang Yang, Minshuo Chen, and Mengdi Wang. Gradient guidance for diffusion models: An optimization perspective, 2024. URL https://arxiv. org/abs/2404.14743, 10

    2024 arXiv

  17. [25]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021

    2021 arXiv

  18. [26]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. URL https://arxiv.org/abs/1706.08500

    2018 arXiv

  19. [27]

    Probability inequalities for sums of bounded random variables

    Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58 0 (301): 0 13--30, 1963 a . ISSN 01621459, 1537274X. URL http://www.jstor.org/stable/2282952

    1963

  20. [28]

    Probability inequalities for sums of bounded random variables

    Wassily Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the American Statistical Association, 58 0 (301): 0 13–30, March 1963 b . ISSN 1537-274X. doi:10.1080/01621459.1963.10500830. URL http://dx.doi.org/10.1080/01621459.1963.10500830

    1963

  21. [29]

    Token merging for training-free semantic binding in text-to-image synthesis

    Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Shahbaz Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, and Yaxing Wang. Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems, 37: 0 137646--...

    2024

  22. [30]

    An optimism-based approach to online evaluation of generative models, 2024 b

    Xiaoyan Hu, Ho fung Leung, and Farzan Farnia. An optimism-based approach to online evaluation of generative models, 2024 b . URL https://arxiv.org/abs/2406.07451

    2024 arXiv

  23. [31]

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

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URL https://...

    2023

  24. [32]

    An information-theoretic evaluation of generative models in learning multi-modal distributions

    Mohammad Jalali, Cheuk Ting Li, and Farzan Farnia. An information-theoretic evaluation of generative models in learning multi-modal distributions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=PdZhf6PiAb

    2023

  25. [33]

    Rethinking fid: Towards a better evaluation metric for image generation, 2024

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation, 2024. URL https://arxiv.org/abs/2401.09603

    2024 arXiv

  26. [34]

    What's ``up'' with vision-language models? investigating their struggle with spatial reasoning

    Amita Kamath, Jack Hessel, and Kai-Wei Chang. What's ``up'' with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, 2023

    2023

  27. [35]

    If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection

    Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. arXiv preprint arXiv:2305.13308, 2023

    2023 arXiv

  28. [36]

    Beyondscene: Higher-resolution human-centric scene generation with pretrained diffusion, 2024

    Gwanghyun Kim, Hayeon Kim, Hoigi Seo, Dong Un Kang, and Se Young Chun. Beyondscene: Higher-resolution human-centric scene generation with pretrained diffusion, 2024. URL https://arxiv.org/abs/2404.04544

    2024 arXiv

  29. [37]

    Dense text-to-image generation with attention modulation

    Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 7701--7711, 2023

    2023

  30. [38]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything, 2023. URL https://arxiv.org/abs/2304.02643

    2023 arXiv

  31. [39]

    Improved precision and recall metric for assessing generative models, 2019

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models, 2019. URL https://arxiv.org/abs/1904.06991

    2019 arXiv

  32. [40]

    Laion-coco 600m

    LAION. Laion-coco 600m. https://laion.ai/blog/laion-coco, 2022

    2022

  33. [41]

    Lee, Taehoon Yoon, and Minhyuk Sung

    Phillip Y. Lee, Taehoon Yoon, and Minhyuk Sung. Groundit: Grounding diffusion transformers via noisy patch transplantation, 2024. URL https://arxiv.org/abs/2410.20474

    2024 arXiv

  34. [42]

    Gligen: Open-set grounded text-to-image generation, 2023 a

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation, 2023 a . URL https://arxiv.org/abs/2301.07093

    2023 arXiv

  35. [43]

    Divide & bind your attention for improved generative semantic nursing

    Yumeng Li, Margret Keuper, Dan Zhang, and Anna Khoreva. Divide & bind your attention for improved generative semantic nursing. In 34th British Machine Vision Conference 2023, BMVC 2023 , 2023 b

    2023

  36. [44]

    Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models, 2024

    Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models, 2024. URL https://arxiv.org/abs/2305.13655

    2024 arXiv

  37. [45]

    Lawrence Zitnick, and Piotr Dollár

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár. Microsoft coco: Common objects in context, 2015. URL https://arxiv.org/abs/1405.0312

    2015 arXiv

  38. [46]

    Evaluating text-to-visual generation with image-to-text generation, 2024

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation, 2024. URL https://arxiv.org/abs/2404.01291

    2024 arXiv

  39. [47]

    Tenenbaum

    Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models, 2023. URL https://arxiv.org/abs/2206.01714

    2023 arXiv

  40. [48]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pp.\ 38--55. Springer, 2024 a

    2024

  41. [49]

    Correcting diffusion generation through resampling

    Yujian Liu, Yang Zhang, Tommi Jaakkola, and Shiyu Chang. Correcting diffusion generation through resampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8713--8723, 2024 b

    2024

  42. [50]

    Inference-time scaling for diffusion models beyond scaling denoising steps

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

    2025 arXiv

  43. [51]

    Diffusion beats autoregressive: An evaluation of compositional generation in text-to-image models

    Arash Marioriyad, Parham Rezaei, Mahdieh Soleymani Baghshah, and Mohammad Hossein Rohban. Diffusion beats autoregressive: An evaluation of compositional generation in text-to-image models. arXiv preprint arXiv:2410.22775, 2024

    arXiv

  44. [52]

    Scaling open-vocabulary object detection

    Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023

    2023

  45. [53]

    Simple open-vocabulary object detection with vision transformers, 2022

    Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple open-vocabulary object detection with vision...

    2022 arXiv

  46. [54]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2022. URL https://arxiv.org/abs/2112.10741

    2022 arXiv

  47. [55]

    Gpt-4 technical report, 2024

    OpenAI. Gpt-4 technical report, 2024. URL https://arxiv.org/abs/2303.08774

    2024 arXiv

  48. [56]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  49. [57]

    Grounded text-to-image synthesis with attention refocusing, 2023

    Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing, 2023. URL https://arxiv.org/abs/2306.05427

    2023 arXiv

  50. [58]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. URL https://arxiv.org/abs/2307.01952

    2023 arXiv

  51. [59]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Con...

    2021

  52. [60]

    Zero-shot text-to-image generation

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

    2021

  53. [61]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. URL https://api.semanticscholar.org/CorpusID:248097655

    2022 arXiv

  54. [62]

    Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment

    Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. ...

    2023

  55. [63]

    Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov

    Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, V.Ya. Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. In Conference on...

    2023

  56. [64]

    Grounded sam: Assembling open-world models for diverse visual tasks, 2024

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. ...

    2024 arXiv

  57. [65]

    Be more diverse than the most diverse: Optimal mixtures of generative models via mixture- UCB bandit algorithms

    Parham Rezaei, Farzan Farnia, and Cheuk Ting Li. Be more diverse than the most diverse: Optimal mixtures of generative models via mixture- UCB bandit algorithms. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=2...

    2025

  58. [66]

    Some aspects of the sequential design of experiments

    Herbert Robbins. Some aspects of the sequential design of experiments. Bulletin of the American Mathematical Society, 58 0 (5): 0 527–535, 1952. ISSN 1088-9485. doi:10.1090/s0002-9904-1952-09620-8. URL http://dx.doi.org/10.1090/S0002-9904-1952-09620-8

    1952 doi

  59. [67]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. 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

  60. [68]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 22500-...

    2023

  61. [69]

    Predicated diffusion: Predicate logic-based attention guidance for text-to-image diffusion models

    Kota Sueyoshi and Takashi Matsubara. Predicated diffusion: Predicate logic-based attention guidance for text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8651--8660, 2024

    2024

  62. [70]

    Instancediffusion: Instance-level control for image generation, 2024

    Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. Instancediffusion: Instance-level control for image generation, 2024. URL https://arxiv.org/abs/2402.03290

    2024 arXiv

  63. [71]

    Wolfe and Robert V

    Douglas A. Wolfe and Robert V. Hogg. On constructing statistics and reporting data. The American Statistician, 25 0 (4): 0 27--30, 1971. ISSN 00031305, 15372731. URL http://www.jstor.org/stable/2682922

    1971

  64. [72]

    R&b: Region and boundary aware zero-shot grounded text-to-image generation, 2023

    Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation, 2023. URL https://arxiv.org/abs/2310.08872

    2023 arXiv

  65. [73]

    Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion, 2023

    Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion, 2023. URL https://arxiv.org/abs/2307.10816

    2023 arXiv

  66. [74]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.), Advances in Neura...

    2023

  67. [75]

    Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms, 2024

    Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms, 2024. URL https://arxiv.org/abs/2401.11708

    2024 arXiv

  68. [76]

    Reco: Region-controlled text-to-image generation

    Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 14246--14255, 2023

    2023

  69. [77]

    Iterative object count optimization for text-to-image diffusion models

    Oz Zafar, Lior Wolf, and Idan Schwartz. Iterative object count optimization for text-to-image diffusion models. arXiv preprint arXiv:2408.11721, 2024

    2024 arXiv

  70. [78]

    Understanding and mitigating compositional issues in text-to-image generative models

    Arman Zarei, Keivan Rezaei, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, and Soheil Feizi. Understanding and mitigating compositional issues in text-to-image generative models. arXiv preprint arXiv:2406.07844, 2024

    2024 arXiv

  71. [79]

    Multi-grained vision language pre-training: Aligning texts with visual concepts

    Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In International Conference on Machine Learning, 2021. URL https://api.semanticscholar.org/CorpusID:244129883

    2021

  72. [80]

    Adding conditional control to text-to-image diffusion models, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. URL https://arxiv.org/abs/2302.05543

    2023 arXiv

  73. [81]

    Realcompo: Balancing realism and compositionality improves text-to-image diffusion models

    Xinchen Zhang, Ling Yang, Yaqi Cai, Zhaochen Yu, Kaini Wang, Jiake Xie, Ye Tian, Minkai Xu, Yong Tang, Yujiu Yang, and Bin Cui. Realcompo: Balancing realism and compositionality improves text-to-image diffusion models. Advances in Neural Information Processing Systems, 2024 a

    2024

  74. [82]

    Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models

    Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. arXiv preprint arXiv:2403.06381, 2024 b

    2024 arXiv

  75. [83]

    Layoutdiffusion: Controllable diffusion model for layout-to-image generation, 2024

    Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation, 2024. URL https://arxiv.org/abs/2303.17189

    2024 arXiv

  76. [84]

    a henb \

    Xingyi Zhou, Vladlen Koltun, and Philipp Kr \"a henb \"u hl. Simple multi-dataset detection. In CVPR, 2022

    2022

  77. [85]

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