pith. sign in

arxiv: 2606.10653 · v1 · pith:NVTLF7R3new · submitted 2026-06-09 · 💻 cs.CV

STEDiff: Strengthening Text Embedding for Text-to-Image Alignment in Diffusion Model

Hailan Zhang , Haipeng Liu , Bo Fu , Yang Wang This is my paper

Pith reviewed 2026-06-27 13:49 UTC · model grok-4.3

classification 💻 cs.CV
keywords text-to-image generationdiffusion modelssemantic alignmenttext embeddingstraining-free methodsemantic enhancement lossT2I-CompBench
0
0 comments X

The pith

A training-free method strengthens text embeddings to improve semantic alignment in diffusion-based image generation.

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

The paper proposes a way to make text-to-image models better follow complex prompts by adjusting the text embeddings directly. It focuses on using the [EOT] token to boost the meaning of parts of the sentence and then uses a loss function to make sure each object ends up in the right place in the image. This matters because current models often miss objects or mix up which features belong to which things, and the method avoids the need for retraining the whole model or providing extra layout information. Evaluations on T2I-CompBench show gains in consistency for intricate cases.

Core claim

STEDiff enhances semantic representations in the text-embedding space by leveraging the [EOT] token to strengthen the relevant semantics of sub-sentences and replacing the corresponding tokens in the original prompt, while incorporating a novel semantic enhancement loss to enforce spatial constraints that map each entity's semantics to its respective image region.

What carries the argument

The [EOT] token used to strengthen sub-sentence semantics in the text embedding, combined with token replacement and a semantic enhancement loss for spatial constraints.

If this is right

  • The method notably improves semantic consistency and generation integrity in complex scenarios on T2I-CompBench.
  • It serves as a computationally efficient alternative to fine-tuning or layout priors.
  • The spatial constraints in the loss ensure precise mapping of entity semantics to image regions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The embedding adjustment could integrate into existing diffusion pipelines for handling more intricate prompts at inference time.
  • The technique might apply to other text-conditioned generation tasks beyond images.
  • Testing on additional benchmarks could reveal whether the approach generalizes across different model architectures.

Load-bearing premise

Directly strengthening sub-sentence semantics via the [EOT] token and applying the semantic enhancement loss will produce precise entity-to-region mappings in the generated image without introducing new artifacts or requiring any model adaptation.

What would settle it

If images generated from complex prompts on T2I-CompBench still show missing objects or incorrect attribute bindings after applying the STEDiff embedding changes, the claim of improved alignment would be falsified.

Figures

Figures reproduced from arXiv: 2606.10653 by Bo Fu, Hailan Zhang, Haipeng Liu, Yang Wang.

Figure 1
Figure 1. Figure 1: Comparison of Different Methods. Facing complex prompts, T2I models often suffer from semantic binding issues. For example, attributes associated with “woman” may be incorrectly bound to “bicycle”, or the “man” entity may fail to be generated correctly. To address these challenges, we propose the STEDiff method. constraints, such as layout guidance [21], [22], or incorporate Large Language Models (LLMs) [2… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of STEDiff. Our method starts by splitting complex prompts and treating each sub-sentence as a clean, prompt-level supervision signal. During the denoising process, we apply STEDiff to the resulting sub-sentences to obtain enhanced embeddings for improved image feature representation. In the early stages of denoising, the Lbind and Lent are used in tandem to update and replace tokens between the o… view at source ↗
Figure 3
Figure 3. Figure 3: Attention visualization. The cross-attention map for each token of the prompt “a man riding a bicycle and a woman walking a dog” and its sub-sentences is visualized. When faced with a complex prompt, there is a noticeable entanglement between different subjects, whereas simple sub￾sentences do not exhibit this behavior. C. Analysis of [EOT] tokens To further reveal the issues that arise when the prompt con… view at source ↗
Figure 5
Figure 5. Figure 5: Analysis of [EOT] tokens. (a)Earlier tokens exhibit higher infor￾mation entropy, which sharply decreases upon reaching the first [EOT] token as the token index increases. (b)When faced with complex prompts, the text embeddings are more likely to deviate from the true semantic space. Lbind = X K k=1 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of image features. (a)Image features are extracted using DINOv3 and visualized through PCA reduction. Compared to the SDXL, our method demonstrates more diverse and rich image features, which contribute to enhanced semantic binding. (b)Statistics of the average feature area calculated from the dimensionality-reduced areas demonstrate that our method reveals more image features. To further valida… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of STEDiff against semantic binding baselines in different scenarios. Whether handling objects and their attributes(second row) or sub-objects(first row), STEDiff consistently maintains high alignment. Moreover, when faced with complex prompts, it reliably generates images that align accurately with the text description. superior semantic binding performance in both object binding an… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study of different optimization terms in the attention mechanism. Cross-attention visualization of the first [EOT] token shows that using any single optimization term alone does not yield ideal results. prompt, we generate one image under each configuration to enable a controlled comparison. Since our method performs activation and replacement by leveraging information from the [EOT] token, visual… view at source ↗
read the original abstract

Although pretrained text-to-image (T2I) generation models can produce high-quality images, they often fail to faithfully reflect the semantic intent of complex prompts due to stochastic noise and inherent model limitations. This issue frequently manifests as the model overlooking specific objects or failing to correctly bind attributes to their corresponding entities, a challenge referred to as semantic alignment. Unlike existing approaches that rely on computationally expensive fine-tuning or labor-intensive layout priors, we propose STEDiff, a training-free method designed to enhance semantic representations directly within the text-embedding space. Specifically, we introduce a method that primarily leverages the [EOT] token to strengthen the relevant semantics of sub-sentences and then replaces the corresponding tokens in the original prompt. Furthermore, a novel semantic enhancement loss is incorporated to enforce spatial constraints, ensuring that the semantics of each entity are precisely mapped to their respective image regions. Extensive quantitative and qualitative evaluations on the T2I-CompBench demonstrate that our method notably improves semantic consistency and generation integrity in complex scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes STEDiff, a training-free method to improve semantic alignment in pretrained text-to-image diffusion models for complex prompts. It strengthens sub-sentence semantics by leveraging the [EOT] token to enhance relevant embeddings and replaces tokens in the original prompt, then applies a novel semantic enhancement loss to enforce spatial constraints that map each entity's semantics to corresponding image regions. Quantitative and qualitative results on T2I-CompBench are reported to show gains in semantic consistency and generation integrity.

Significance. A training-free embedding-space intervention that reliably produces precise entity-to-region bindings would be a useful, low-cost addition to the T2I alignment literature. However, the central mechanism (how a purely token-level loss translates semantic signals into spatially accurate denoising trajectories without explicit cross-attention or image-space terms) is not shown to be load-bearing or robust; if the loss reduces to implicit model behavior, the claimed improvements would not generalize beyond the evaluated benchmark.

major comments (1)
  1. [Abstract] Abstract: the claim that the semantic enhancement loss 'ensures that the semantics of each entity are precisely mapped to their respective image regions' is not supported by any described dependence on cross-attention maps, positional encodings, or image-space terms. Without such dependence, the loss cannot be guaranteed to produce the required spatial bindings from embedding modifications alone.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the semantic enhancement loss 'ensures that the semantics of each entity are precisely mapped to their respective image regions' is not supported by any described dependence on cross-attention maps, positional encodings, or image-space terms. Without such dependence, the loss cannot be guaranteed to produce the required spatial bindings from embedding modifications alone.

    Authors: We agree that the abstract wording overstates the mechanism. The semantic enhancement loss is applied to the strengthened embeddings to encourage spatial consistency, but the manuscript does not describe explicit dependence on cross-attention maps, positional encodings, or image-space terms; any spatial effect arises implicitly through the diffusion process. We will revise the abstract to replace 'ensures that the semantics of each entity are precisely mapped' with 'encourages improved mapping of entity semantics to image regions' and will similarly qualify the claim in the main text. The reported gains on T2I-CompBench remain as empirical evidence. revision: yes

Circularity Check

0 steps flagged

No circularity: training-free embedding edits and external benchmark evaluation remain independent of fitted inputs or self-citation chains

full rationale

The paper describes a training-free procedure that strengthens sub-sentence semantics via the [EOT] token, token replacement, and a novel semantic enhancement loss to enforce spatial constraints in embedding space. No equations or steps reduce a claimed prediction to a fitted parameter by construction, no load-bearing self-citations are invoked to justify uniqueness or ansatzes, and the quantitative results are reported against the external T2I-CompBench benchmark rather than internal data splits. The derivation chain therefore consists of explicit, externally verifiable modifications to existing diffusion-model components without tautological closure.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; the approach is described as operating on existing diffusion-model components and the [EOT] token.

pith-pipeline@v0.9.1-grok · 5710 in / 1167 out tokens · 38310 ms · 2026-06-27T13:49:09.017327+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 6 linked inside Pith

  1. [1]

    ediff-i: Text-to-image diffu- sion models with an ensemble of expert denoisers,

    Y . Balaji, S. Nah, X. Huang, A. Vahdat, J. Song, Q. Zhang, K. Kreis, M. Aittala, T. Aila, S. Laine,et al., “ediff-i: Text-to-image diffu- sion models with an ensemble of expert denoisers,”arXiv preprint arXiv:2211.01324, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Zero-shot text-to-image generation,

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” inInternational conference on machine learning, pp. 8821–8831, Pmlr, 2021

    2021

  3. [3]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al., “Photorealistic text-to-image diffusion models with deep language understanding,”Advances in neural information processing systems, vol. 35, pp. 36479–36494, 2022

    2022

  4. [4]

    Hierarchical text-conditional image generation with clip latents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,”arXiv preprint arXiv:2204.06125, vol. 1, no. 2, p. 3, 2022

    Pith/arXiv arXiv 2022

  5. [5]

    One stone with two birds: A null-text- null frequency-aware diffusion models for text-guided image inpainting,

    H. Liu, Y . Wang, and M. Wang, “One stone with two birds: A null-text- null frequency-aware diffusion models for text-guided image inpainting,” inThe Thirty-ninth Annual Conference on Neural Information Process- ing Systems, 2025

    2025

  6. [6]

    Structure matters: Tackling the semantic discrepancy in diffusion models for image inpaint- ing,

    H. Liu, Y . Wang, B. Qian, M. Wang, and Y . Rui, “Structure matters: Tackling the semantic discrepancy in diffusion models for image inpaint- ing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8038–8047, 2024

    2024

  7. [7]

    Progressive learning with multi-scale attention network for cross-domain vehicle re-identification,

    Y . Wang, J. Peng, H. Wang, and M. Wang, “Progressive learning with multi-scale attention network for cross-domain vehicle re-identification,” Science China Information Sciences, vol. 65, no. 6, p. 160103, 2022

    2022

  8. [8]

    Divide & bind your attention for improved generative semantic nursing,

    Y . Li, M. Keuper, D. Zhang, and A. Khoreva, “Divide & bind your attention for improved generative semantic nursing,”arXiv preprint arXiv:2307.10864, 2023

    arXiv 2023

  9. [9]

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

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,”arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel,et al., “Scaling rectified flow transformers for high-resolution image synthesis,” inForty-first international conference on machine learning, 2024

    2024

  11. [11]

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

    R. Rassin, E. Hirsch, D. Glickman, S. Ravfogel, Y . Goldberg, and G. Chechik, “Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment,”Advances in Neural Information Processing Systems, vol. 36, pp. 3536–3559, 2023

    2023

  12. [12]

    Expressive text-to-image generation with rich text,

    S. Ge, T. Park, J.-Y . Zhu, and J.-B. Huang, “Expressive text-to-image generation with rich text,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7545–7556, 2023

    2023

  13. [13]

    Switchable online knowledge distillation,

    B. Qian, Y . Wang, H. Yin, R. Hong, and M. Wang, “Switchable online knowledge distillation,” inEuropean Conference on Computer Vision, pp. 449–466, Springer, 2022

    2022

  14. [14]

    Adaptive data-free quantiza- tion,

    B. Qian, Y . Wang, R. Hong, and M. Wang, “Adaptive data-free quantiza- tion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7960–7968, 2023

    2023

  15. [15]

    Unpacking the gap box against data-free knowledge distillation,

    Y . Wang, B. Qian, H. Liu, Y . Rui, and M. Wang, “Unpacking the gap box against data-free knowledge distillation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 9, pp. 6280–6291, 2024

    2024

  16. [16]

    Comat: Aligning text-to-image diffusion model with image- to-text concept matching,

    D. Jiang, G. Song, X. Wu, R. Zhang, D. Shen, Z. Zong, Y . Liu, and H. Li, “Comat: Aligning text-to-image diffusion model with image- to-text concept matching,”Advances in Neural Information Processing Systems, vol. 37, pp. 76177–76209, 2024

    2024

  17. [17]

    Ella: Equip diffusion models with llm for enhanced semantic alignment,

    X. Hu, R. Wang, Y . Fang, B. Fu, P. Cheng, and G. Yu, “Ella: Equip diffusion models with llm for enhanced semantic alignment,”arXiv preprint arXiv:2403.05135, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Ranni: Taming text-to-image diffusion for accurate instruction following,

    Y . Feng, B. Gong, D. Chen, Y . Shen, Y . Liu, and J. Zhou, “Ranni: Taming text-to-image diffusion for accurate instruction following,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4744–4753, 2024

    2024

  19. [19]

    Few- shot referring video single-and multi-object segmentation via cross- modal affinity with instance sequence matching,

    H. Liu, G. Li, M. Gao, X. Zhen, F. Zheng, and Y . Wang, “Few- shot referring video single-and multi-object segmentation via cross- modal affinity with instance sequence matching,”International Journal of Computer Vision, vol. 133, no. 8, pp. 5610–5628, 2025

    2025

  20. [20]

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

    J. Chen, C. Ge, E. Xie, Y . Wu, L. Yao, X. Ren, Z. Wang, P. Luo, H. Lu, and Z. Li, “Pixart-σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation,” inEuropean Conference on Computer Vision, pp. 74–91, Springer, 2024

    2024

  21. [21]

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

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

    2023

  22. [22]

    Grounded text-to-image synthesis with attention refocusing,

    Q. Phung, S. Ge, and J.-B. Huang, “Grounded text-to-image synthesis with attention refocusing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7932–7942, 2024

    2024

  23. [23]

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

    L. Lian, B. Li, A. Yala, and T. Darrell, “Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models,”arXiv preprint arXiv:2305.13655, 2023

    arXiv 2023

  24. [24]

    Llm blueprint: Enabling text-to-image generation with complex and detailed prompts,

    H. Gani, S. F. Bhat, M. Naseer, S. Khan, and P. Wonka, “Llm blueprint: Enabling text-to-image generation with complex and detailed prompts,” arXiv preprint arXiv:2310.10640, 2023

    arXiv 2023

  25. [25]

    Rethinking data-free quan- tization as a zero-sum game,

    B. Qian, Y . Wang, R. Hong, and M. Wang, “Rethinking data-free quan- tization as a zero-sum game,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, pp. 9489–9497, 2023

    2023

  26. [26]

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

    T. Hu, L. Li, J. van de Weijer, H. Gao, F. Shahbaz Khan, J. Yang, M.-M. Cheng, K. Wang, and Y . Wang, “Token merging for training- free semantic binding in text-to-image synthesis,”Advances in Neural Information Processing Systems, vol. 37, pp. 137646–137672, 2024

    2024

  27. [27]

    A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization,

    C.-Y . Chen, C. Tseng, L.-W. Tsao, and H.-H. Shuai, “A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization,”Advances in Neural Information Processing Systems, vol. 37, pp. 57944–57969, 2024

    2024

  28. [28]

    Delving globally into texture and structure for image inpainting,

    H. Liu, Y . Wang, M. Wang, and Y . Rui, “Delving globally into texture and structure for image inpainting,” inProceedings of the 30th ACM International Conference on Multimedia, pp. 1270–1278, 2022

    2022

  29. [29]

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

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

    2023

  30. [30]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684–10695, 2022

    2022

  31. [31]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  32. [32]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    B. F. Labs, S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esser, S. Kulal, K. Lacey, Y . Levi, C. Li, D. Lorenz, J. M ¨uller, D. Podell, R. Rombach, H. Saini, A. Sauer, and L. Smith, “Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,” 2025

    2025

  33. [33]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  34. [34]

    Learning transferable visual models from natural language supervision,

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

    2021

  35. [35]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  36. [36]

    Be yourself: Bounded attention for multi-subject text-to-image generation,

    O. Dahary, O. Patashnik, K. Aberman, and D. Cohen-Or, “Be yourself: Bounded attention for multi-subject text-to-image generation,” inEuro- pean Conference on Computer Vision, pp. 432–448, Springer, 2024

    2024

  37. [37]

    Phased consistency models,

    F.-Y . Wang, Z. Huang, A. Bergman, D. Shen, P. Gao, M. Lingelbach, K. Sun, W. Bian, G. Song, Y . Liu,et al., “Phased consistency models,” Advances in neural information processing systems, vol. 37, pp. 83951– 84009, 2024

    2024

  38. [38]

    Rethinking the spa- tial inconsistency in classifier-free diffusion guidance,

    D. Shen, G. Song, Z. Xue, F.-Y . Wang, and Y . Liu, “Rethinking the spa- tial inconsistency in classifier-free diffusion guidance,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9370–9379, 2024

    2024

  39. [39]

    Plug-and-play diffu- sion features for text-driven image-to-image translation,

    N. Tumanyan, M. Geyer, S. Bagon, and T. Dekel, “Plug-and-play diffu- sion features for text-driven image-to-image translation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1921–1930, 2023

    1921

  40. [40]

    Compositional text-to-image synthesis with attention map control of diffusion models,

    R. Wang, Z. Chen, C. Chen, J. Ma, H. Lu, and X. Lin, “Compositional text-to-image synthesis with attention map control of diffusion models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 5544–5552, 2024

    2024

  41. [41]

    Instancediffusion: Instance-level control for image generation,

    X. Wang, T. Darrell, S. S. Rambhatla, R. Girdhar, and I. Misra, “Instancediffusion: Instance-level control for image generation,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 6232–6242, 2024

    2024

  42. [42]

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

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

    2024

  43. [43]

    Mulan: Multimodal-llm agent for progressive and interactive multi-object diffu- sion,

    S. Li, R. Wang, C.-J. Hsieh, M. Cheng, and T. Zhou, “Mulan: Multimodal-llm agent for progressive and interactive multi-object diffu- sion,”arXiv preprint arXiv:2402.12741, 2024

    arXiv 2024

  44. [44]

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

    H. Chefer, Y . Alaluf, Y . Vinker, L. Wolf, and D. Cohen-Or, “Attend-and- excite: Attention-based semantic guidance for text-to-image diffusion models,”ACM transactions on Graphics (TOG), vol. 42, no. 4, pp. 1– 10, 2023

    2023

  45. [45]

    Geometrical properties of text token embeddings for strong semantic binding in text- to-image generation,

    H. Seo, J. Bang, H. Lee, J. Lee, B. H. Lee, and S. Y . Chun, “Geometrical properties of text token embeddings for strong semantic binding in text- to-image generation,”arXiv preprint arXiv:2503.23011, 2025

    arXiv 2025

  46. [46]

    Uncovering the text embedding in text-to-image diffusion models,

    H. Yu, H. Luo, F. Wang, and F. Zhao, “Uncovering the text embedding in text-to-image diffusion models,”arXiv preprint arXiv:2404.01154, 2024

    arXiv 2024

  47. [47]

    Core: Context-regularized text embedding learning for text-to-image personalization,

    F. Wu, Y . Pang, J. Zhang, L. Pang, J. Yin, B. Zhao, Q. Li, and X. Mao, “Core: Context-regularized text embedding learning for text-to-image personalization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 8377–8385, 2025

    2025

  48. [48]

    Get what you want, not what you don’t: Image content suppression for text-to-image diffusion models,

    S. Li, J. van de Weijer, T. Hu, F. S. Khan, Q. Hou, Y . Wang, and J. Yang, “Get what you want, not what you don’t: Image content suppression for text-to-image diffusion models,”arXiv preprint arXiv:2402.05375, 2024

    arXiv 2024

  49. [49]

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

    W. Feng, X. He, T.-J. Fu, V . Jampani, A. Akula, P. Narayana, S. Basu, X. E. Wang, and W. Y . Wang, “Training-free structured diffusion guidance for compositional text-to-image synthesis,”arXiv preprint arXiv:2212.05032, 2022

    arXiv 2022

  50. [50]

    Weighted nuclear norm minimization with application to image denoising,

    S. Gu, L. Zhang, W. Zuo, and X. Feng, “Weighted nuclear norm minimization with application to image denoising,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 2862– 2869, 2014

    2014

  51. [51]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  52. [52]

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

    J. Xu, X. Liu, Y . Wu, Y . Tong, Q. Li, M. Ding, J. Tang, and Y . Dong, “Imagereward: Learning and evaluating human preferences for text-to- image generation,” 2023

    2023

  53. [53]

    Detail++: Training-free detail enhancer for text-to-image diffusion models,

    L. Chen, J. Wang, Z. Pan, B. Zhu, X. Yang, and C. Zhang, “Detail++: Training-free detail enhancer for text-to-image diffusion models,”arXiv preprint arXiv:2507.17853, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    Enhanc- ing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models,

    Y . Zhang, T. T. Tzun, L. W. Hern, T. Sim, and K. Kawaguchi, “Enhanc- ing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models,” 2024

    2024

  55. [55]

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

    R. Rassin, E. Hirsch, D. Glickman, S. Ravfogel, Y . Goldberg, and G. Chechik, “Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment,” 2024

    2024

  56. [56]

    spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing,

    M. Honnibal, “spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing,” (No Title), 2017

    2017

  57. [57]

    Play- ground v2

    D. Li, A. Kamko, A. Sabet, E. Akhgari, L. Xu, and S. Doshi, “Play- ground v2.”