pith. sign in

arxiv: 2606.03715 · v1 · pith:VCF7QPFFnew · submitted 2026-06-02 · 💻 cs.CV

Text-to-Image Models Need Less from Text Encoders Than You Think

Nurit Spingarn , Noa Cohen , Tamar Rott Shaham , Tomer Michaeli This is my paper

Pith reviewed 2026-06-28 11:06 UTC · model grok-4.3

classification 💻 cs.CV
keywords text-to-image generationdiffusion transformerstext embeddingsword orderpositional embeddingscontextual informationprompt conditioning
0
0 comments X

The pith

Text-to-image diffusion transformer models generate high-quality images guided only by individual word meanings and their order.

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

The paper shows that standard text-to-image models based on diffusion transformers succeed at producing images that match prompt descriptions even when fed a stripped-down text embedding containing nothing beyond the meanings of separate words and the sequence in which they appear. The authors build this simpler representation by taking the text encoder's output, collapsing multi-token words into single units, and retaining only the positional signals that mark word order while discarding any cross-word context. A reader would care because the result challenges the widespread view that rich contextual signals such as attribute binding or compositional structure must be supplied by the text encoder for the model to work well. Instead the finding points to the image-generation network itself as the component that interprets and assembles those structures from the simpler input.

Core claim

Text-to-image diffusion transformer-based models commonly rely only on two relatively straightforward aspects of text representations: the merging of adjacent tokens into a word representation for words spanning multiple tokens, and word order, which is imprinted by the positional embedding of the text-encoder. A new text embedding that encodes only individual word meanings and order but lacks any contextual information about the full prompt is sufficient to guide image generation, achieving visual quality and text fidelity on par with full text embedding-guided generation. This demonstrates that the models often do not use the rich information encoded in the text embedding beyond individual

What carries the argument

The bag of position-tagged words representation, which collapses multi-token words and keeps only positional order signals while removing all cross-word contextual information from the prompt.

If this is right

  • The image-generation network itself performs the decoding of complex linguistic structures such as compositionality and attribute binding.
  • Text encoders do not need to supply contextual information across the full prompt for effective conditioning of these models.
  • Merging adjacent tokens into word units and preserving word order are the only text-encoder features required for competitive generation quality.
  • Simpler, context-free text representations can replace richer embeddings without measurable loss in visual quality or prompt fidelity.

Where Pith is reading between the lines

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

  • Designers could explore training or fine-tuning image models on even lighter text inputs to reduce encoder complexity.
  • The result raises the question of whether similar minimal representations would suffice in other conditional generation settings such as text-to-video.
  • Interpretability work could now isolate which layers inside the image model are responsible for reassembling word order into scene structure.
  • Prompt engineering might shift focus toward ensuring clear word sequences rather than crafting elaborate contextual phrasing.

Load-bearing premise

The constructed bag-of-position-tagged-words embedding truly encodes only individual word meanings and order but lacks any contextual information about the full prompt.

What would settle it

Running the same set of complex prompts through both the full text embedding and the bag-of-position-tagged-words embedding and finding that the latter produces visibly poorer attribute binding or compositional accuracy would falsify the claim.

Figures

Figures reproduced from arXiv: 2606.03715 by Noa Cohen, Nurit Spingarn, Tamar Rott Shaham, Tomer Michaeli.

Figure 1
Figure 1. Figure 1: Contextless text embeddings are often enough. We find that when pretrained TTI models are conditioned on text embeddings that are stripped off of any contextual information, they maintain high visual quality and prompt adherence. This surprising behavior is exhibited even for complex prompts that involve attribute binding, spatial relations, and numeracy. We show that the capability of generating text-adhe… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of our three contextless text embeddings. We propose three embedding types of increasing richness: (i) Bag of Tokens, where each token is represented independently; (ii) Bag of Words, where tokens are merged into word￾level representations; and (iii) Bag of Position￾Tagged Words, where word embeddings addition￾ally reflect their position in the prompt. To answer this question, we construct mod… view at source ↗
Figure 3
Figure 3. Figure 3: Construction of contextless embedding. To understand which types of information in the text embeddings are primarily utilized by the image model, we construct three contextless embeddings. Each begins by tokenizing the prompt (e.g., “a red cube”) into discrete tokens by the text encoder’s tokenizer (e.g., “a”,“red”,“cu”,“be”). These are processed by an eraser module which strips away targeted contextual in… view at source ↗
Figure 4
Figure 4. Figure 4: Visual examples by prompt complexity. The BoT and BoW embeddings provide the image model with sufficient information for relatively simpler cases. The BoPTW embedding can support more complex prompts. All images were generated with FLUX.1 Schnell. such words, we introduce the BoW embedding, which refines the previous approach by preserving the cohesion of multi-token words. Specifically, in BoW, tokens rep… view at source ↗
Figure 5
Figure 5. Figure 5: Image generation with the different contextless embeddings. For complex text prompts, the BoT embeddings do not suffice for generating text-adherent images. While the BoW embeddings sometimes provide sufficient improvement, the combination of word-level tokenization with positional information provided by the BoPTW embeddings, consistently enables generating images that closely adhere to the prompt and are… view at source ↗
Figure 6
Figure 6. Figure 6: Text alignment comparison. Image pairs generated from full versus contextless em￾beddings are compared using Gemma as an automated evaluator. Notably, the BoPTW embedding (bottom row) achieve a non-inferiority rate of at least 65% with respect to the full embedding (the combination of the two greenish areas) for most benchmarks and models. This is while the non￾inferiority rate of the full-embedding with r… view at source ↗
Figure 7
Figure 7. Figure 7: Text alignment across categories. Breakdown on VLM responses by category on images generated with the BoPTW embedding, for the DrawBench and GenEval benchmarks. Results are sorted by the mean non-inferiority rate across all evaluated models to highlight which categories are most resilient to the removal of full prompt context. For the GenEval dataset, we further report in Tab. S4 the task-specific scores a… view at source ↗
Figure 8
Figure 8. Figure 8: Most and least successful categories. The top and bottom pairs of rows show visual examples from the two most successful and two least successful categories, respectively, in the DrawBench and GenEval datasets. Each example compares the image generated from the BoPTW embedding to that generated from the full embedding. DiT vs. U-Net. While the Imagen work [28] highlighted the importance of a dedicated text… view at source ↗
read the original abstract

Text-to-image models rely on text prompts as their primary interface to human intent. Prompts are encoded by a text encoder into embeddings that condition the image generation process. Beyond individual token meanings, text embeddings encode contextual information across the full prompt, such as compositionality and attribute binding. However, whether image models actually exploit this richer information remains underexplored. Here, we address the question: Which aspects of text representation are essential for image generation? We show that text-to-image diffusion transformer-based models commonly rely only on two relatively straightforward aspects of text representations: (i) the merging of adjacent tokens into a word representation, for words spanning multiple tokens, and (ii) word order, which is imprinted by the positional embedding of the text-encoder. To show this, we construct a new text embedding that encodes only individual word meanings and order but lacks any contextual information about the full prompt. We find that this bag of position-tagged words representation is sufficient to successfully guide image generation, achieving visual quality and text fidelity that are on par with full text embedding-guided generation. This demonstrates that, contrary to common belief, text-to-image models often do not use the rich information encoded in the text embedding beyond individual word meanings and word order. Instead, the decoding of complex linguistic structures is performed by the image model itself. Project webpage: https://nsping13.github.io/contextless-TTI/

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 manuscript claims that text-to-image diffusion transformer models rely primarily on two aspects of text representations: merging of adjacent tokens into word representations and word order via positional embeddings. The authors construct a 'bag of position-tagged words' embedding that encodes only individual word meanings and order without full-prompt contextual information (e.g., compositionality or attribute binding from cross-token attention in the text encoder). They report that this reduced representation guides image generation with visual quality and text fidelity on par with full text embeddings, implying that the image model itself performs the decoding of complex linguistic structures.

Significance. If substantiated, the result would meaningfully revise understanding of the division of labor in T2I systems by showing that rich contextual encoding from text encoders is often not exploited by the image model. The explicit construction isolating per-word outputs plus positional information is a methodological strength that directly targets the question of what information is load-bearing.

major comments (1)
  1. [Abstract / Experiments] The central claim of on-par performance is load-bearing for the conclusion yet the provided abstract (and any corresponding results section) supplies no details on evaluation metrics, baselines, statistical significance, number of prompts or images evaluated, or controls for confounding factors such as prompt selection.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment of the work's significance and methodological contribution. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract / Experiments] The central claim of on-par performance is load-bearing for the conclusion yet the provided abstract (and any corresponding results section) supplies no details on evaluation metrics, baselines, statistical significance, number of prompts or images evaluated, or controls for confounding factors such as prompt selection.

    Authors: We agree that the abstract is high-level and does not enumerate the concrete evaluation protocol. The Experiments section of the manuscript does describe the metrics (FID for visual quality and CLIP-based text alignment), the prompt sets used, and direct comparison to the unmodified text-encoder baseline. However, explicit statements of sample size, statistical significance testing, and prompt-selection controls are not as prominently listed as they should be. We will revise both the abstract (to include the key quantitative results and evaluation scope) and the Experiments section (to add a dedicated paragraph on sample sizes, significance testing, and controls) so that the central claim is fully substantiated in the text. revision: yes

Circularity Check

0 steps flagged

Empirical construction is self-contained; no circular reduction

full rationale

The paper's core contribution is an explicit construction of a 'bag of position-tagged words' embedding that isolates per-word meanings plus positional order while removing cross-token context from the text encoder. Performance is then measured empirically against full embeddings on image generation quality. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text that would reduce the sufficiency claim to a definitional tautology. The construction directly targets the weakest assumption and the result follows from the comparison rather than from any internal loop.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract alone.

pith-pipeline@v0.9.1-grok · 5793 in / 980 out tokens · 37082 ms · 2026-06-28T11:06:09.200303+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

38 extracted references · 14 canonical work pages · 7 internal anchors

  1. [1]

    Claude sonnet 4.5

    Anthropic. Claude sonnet 4.5. https://www.anthropic.com/, 2025. Large language model

    2025

  2. [2]

    Demystifying MMD GANs

    Mikołaj Bi´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans.arXiv preprint arXiv:1801.01401, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  3. [3]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558–3568, 2021

    2021

  4. [4]

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models.arXiv preprint arXiv:2301.13826, 2023

    Hila Chefer, Omer Tov, Roni Paiss, Lior Wolf, et al. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models.arXiv preprint arXiv:2301.13826, 2023

    work page arXiv 2023

  5. [5]

    Vision language models learn to assess images with specialists

    Quyet V Do, Seunghyun Yoon, Ruiyi Zhang, Thiloshon Nagarajah, Trung Bui, and Viet Dac Lai. Vision language models learn to assess images with specialists. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1126–1135, 2026

    2026

  6. [6]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024

    2024

  7. [7]

    Geneval: An object-focused framework for evaluating text-to-image alignment.Advances in Neural Information Processing Systems, 36:52132–52152, 2023

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment.Advances in Neural Information Processing Systems, 36:52132–52152, 2023

    2023

  8. [8]

    Prompt-to-Prompt Image Editing with Cross Attention Control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [9]

    Sugarcrepe: Fixing compositionality in vision-language models

    Jack Hessel, Youngjae Yu, Yejin Kwon, and Yejin Choi. Sugarcrepe: Fixing compositionality in vision-language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  10. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30, 2017

    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.Advances in neural information processing systems, 30, 2017

    2017

  11. [11]

    A structural probe for finding syntax in word representations

    John Hewitt and Christopher D Manning. A structural probe for finding syntax in word representations. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138, 2019

    2019

  12. [12]

    What does bert learn about the structure of language? InProceedings of the 57th annual meeting of the association for computational linguistics, pages 3651–3657, 2019

    Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. What does bert learn about the structure of language? InProceedings of the 57th annual meeting of the association for computational linguistics, pages 3651–3657, 2019

    2019

  13. [13]

    Mistral 7B

    Albert Q. Jiang et al. Mistral 7b.arXiv preprint arXiv:2310.06825, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Flux.1: A family of text-to-image models

    Black Forest Labs. Flux.1: A family of text-to-image models. Technical Report, 2024

    2024

  15. [15]

    FLUX.2: Frontier Visual Intelligence

    Black Forest Labs. FLUX.2: Frontier Visual Intelligence. https://bfl.ai/blog/flux-2, 2025

    2025

  16. [16]

    Prometheus-vision: Vision-language model as a judge for fine-grained evaluation

    Seongyun Lee, Seungone Kim, Sue Park, Geewook Kim, and Minjoon Seo. Prometheus-vision: Vision-language model as a judge for fine-grained evaluation. InFindings of the Association for Computational Linguistics: ACL 2024, pages 11286–11315, 2024. 10

    2024

  17. [17]

    Vhelm: A holistic evaluation of vision language models.Advances in Neural Information Processing Systems, 37:140632–140666, 2024

    Tony Lee, Haoqin Tu, Chi H Wong, Wenhao Zheng, Yiyang Zhou, Yifan Mai, Josselin S Roberts, Michihiro Yasunaga, Huaxiu Yao, Cihang Xie, et al. Vhelm: A holistic evaluation of vision language models.Advances in Neural Information Processing Systems, 37:140632–140666, 2024

    2024

  18. [18]

    Deleaker: Improving text-to-image diffusion models via deletion and leakage control.arXiv preprint arXiv:2310.00000, 2023

    Yujin Li et al. Deleaker: Improving text-to-image diffusion models via deletion and leakage control.arXiv preprint arXiv:2310.00000, 2023

    work page arXiv 2023

  19. [19]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean conference on computer vision, pages 740–755. Springer, 2014

    2014

  20. [20]

    Understanding the limitations of clip for compositionality.arXiv preprint arXiv:2305.00000, 2023

    Sahil Palit et al. Understanding the limitations of clip for compositionality.arXiv preprint arXiv:2305.00000, 2023

    work page arXiv 2023

  21. [21]

    Emergence of hidden capabilities: Exploring learning dynamics in concept space.Advances in Neural Information Processing Systems, 37:84698–84729, 2024

    Core F Park, Maya Okawa, Andrew Lee, Hidenori Tanaka, and Ekdeep S Lubana. Emergence of hidden capabilities: Exploring learning dynamics in concept space.Advances in Neural Information Processing Systems, 37:84698–84729, 2024

    2024

  22. [22]

    Dreambench++: A human-aligned benchmark for personalized image generation.arXiv preprint arXiv:2406.16855, 2024

    Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation.arXiv preprint arXiv:2406.16855, 2024

    work page arXiv 2024

  23. [23]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    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.arXiv preprint arXiv:2307.01952, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    Learning transferable visual models from natural language supervision.Proceedings of the International Conference on Machine Learning (ICML), 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision.Proceedings of the International Conference on Machine Learning (ICML), 2021

    2021

  25. [25]

    Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of Machine Learning Research, 21(140):1–67, 2020

    Colin Raffel, Noam Shazeer, Adam Roberts, et al. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of Machine Learning Research, 21(140):1–67, 2020

    2020

  26. [26]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  27. [27]

    High-resolution image synthesis with latent diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, et al. High-resolution image synthesis with latent diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  28. [28]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    2022

  29. [29]

    What the daam: Interpreting stable diffusion using cross attention.arXiv preprint arXiv:2210.04885, 2023

    Raphael Tang, Yixuan Zhang, et al. What the daam: Interpreting stable diffusion using cross attention.arXiv preprint arXiv:2210.04885, 2023

    work page arXiv 2023

  30. [30]

    Gemma 3 Technical Report

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Bert rediscovers the classical nlp pipeline

    Ian Tenney, Dipanjan Das, and Ellie Pavlick. Bert rediscovers the classical nlp pipeline. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 4593–4601, 2019

    2019

  32. [32]

    Girshick

    Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5228–5238, 2022. doi: 10.1109/CVPR52688.2022.00517. 11

    work page doi:10.1109/cvpr52688.2022.00517 2022

  33. [33]

    Diffusion lens: Interpreting text encoders in text-to-image pipelines

    Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, and Yonatan Belinkov. Diffusion lens: Interpreting text encoders in text-to-image pipelines. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), pages 9713–9728, 2024. doi: 10.18653/v1/2024.acl-long.524

    work page doi:10.18653/v1/2024.acl-long.524 2024

  34. [34]

    Circuit Mechanisms for Spatial Relation Generation in Diffusion Transformers

    Binxu Wang, Jingxuan Fan, and Xu Pan. Circuit mechanisms for spatial relation generation in diffusion transformers.arXiv preprint arXiv:2601.06338, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  35. [35]

    Scaling down text encoders of text-to- image diffusion models

    Lifu Wang, Daqing Liu, Xinchen Liu, and Xiaodong He. Scaling down text encoders of text-to- image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18424–18433, 2025

    2025

  36. [36]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    When and why vision-language models behave like bags-of-words, and what to do about it

    Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it. In International Conference on Learning Representations (ICLR), 2023

    2023

  38. [38]

    a red cube

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in neural information processing systems, 36:46595–46623, 2023. 12 Appendix A Additional Results A.1 Encoding positional information in the text embedding T...

    2023