Linguistic Binding in Diffusion Models: Enhancing Attribute Correspondence through Attention Map Alignment

Text-conditioned image generation models often generate incorrect associations between entities and their visual attributes. This reflects an impaired mapping between linguistic binding of entities and modifiers in the prompt and visual binding of the corresponding elements in the generated image. As one notable example, a query like "a pink sunflower and a yellow flamingo" may incorrectly produce an image of a yellow sunflower and a pink flamingo. To remedy this issue, we propose SynGen, an approach which first syntactically analyses the prompt to identify entities and their modifiers, and then uses a novel loss function that encourages the cross-attention maps to agree with the linguistic binding reflected by the syntax. Specifically, we encourage large overlap between attention maps of entities and their modifiers, and small overlap with other entities and modifier words. The loss is optimized during inference, without retraining or fine-tuning the model. Human evaluation on three datasets, including one new and challenging set, demonstrate significant improvements of SynGen compared with current state of the art methods. This work highlights how making use of sentence structure during inference can efficiently and substantially improve the faithfulness of text-to-image generation.

Paper

References (24)

Scroll for more · 12 remaining

Similar papers

Peer review

Reviewer LWw27/10 · confidence 3/52023-07-01

Summary

This paper presents an approach called “SynGen” to improve attribute binding between nouns and their modifiers for text-to-image latent diffusion models, specifically Stable Diffusion. They propose a two-part loss on attention maps that (1) encourages attention maps of nouns and their modifiers to be similar, and (2) encourages attention maps between the noun/modifier to be different from those of other tokens in the prompt. Datasets and baselines: They run experiments on three datasets: the attribute binding contrast set (ABC-6K) from the Structured Diffusion (baseline 1) paper, data from the Attend-and-Excite (baseline 2), and a newly proposed challenge set called Diverse Visual Modifier Prompts. Evaluation: (1) Human evals for concept separation and visual appeal. They find that SynGen outperforms baselines. (2) Qualitative analysis with sample prompts and generations for Syngen and the two baselines, demonstrating that SynGen overcomes failure cases of the other two models. (3) Ablations on the loss components and loss weighting.

Strengths

1. This paper tackles an important and challenging drawback of text-to-image diffusion models struggling to faithfully generate objects with the right attributes. 2. This paper proposes an interesting two-part loss that imposes localized pairwise constraints on the attention maps between image patches and token embeddings. The first constraint enforces that the same image patches attend to both the noun and its modifiers. The second constraint enforces that image patches attending to these nouns/modifiers do not attend to any other tokens in the prompt. 3. This paper attempts to evaluate on a large set of prompts including those from prior work as well as proposing their own challenge set. 4. This paper attempts to better understand the effects and side-effects of the proposed losses via ablations.

Weaknesses

1. The proposed method appears to be strongly grounded in the framework of the Attend-and-Excite paper. The writing doesn’t highlight this connection while introducing the approach in Section 2, beyond a minor footnote. Section 2 would probably need some revision to better draw this connection. 2. Considering that the paper is tackling a very specific problem, the general “content separation” rating collected in the human eval seems pretty weak. Why not collect finer-grained scores for: number of objects in the prompt, number of objects from prompt generated, number of objects from prompt generated with the correct attributes? Wouldn’t this give a much better sense of task success and could potentially allow computing recall? 3. The loss function ablation mentions that sometimes objects are omitted from the generated image. Is combining the proposed losses with global constraint defined in the Attend-and-Excite paper a viable option? Since the framework is identical, wouldn’t combining the losses have mitigated this issue? The Attend-and-Excite paper talks about attribute binding as well, so this appears to be a variant to try. Nits: 1. Maybe more explicitly state that this is an inference-only guidance procedure? 2. There’s a typo on line 112 for the equation: $\\nabla_{z_t}\\mathcal{L}$ not $\\nabla z_t\\mathcal{L}$. Also $z’_t$ not $z’t$.

Questions

1. Was there a reason to not include numerical modifiers in the study? 2. Are there unexpected side-effects of such constraints on model generalization?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

One limitation that was perhaps overlooked is the lack of standardized evaluation. Having to rely on different crowdsourcing tasks and qualitative comparisons makes it much harder to track progress.

Reviewer EHJQ8/10 · confidence 3/52023-07-06

Summary

A frequent issue in text-prompted image generation (and in many other grounded language scenarios) is that a model will treat text akin to a bag of words and ignore syntactic relationships, such as which adjective attaches to which noun -- this is referred to as a problem with lingusitic binding. The paper proposes addressing this problem in diffusion models by adding an extra step to the diffusion process that nudges the image in the process of being generated so as to respect modifier relationships extracted from a syntactic parse of the prompt. This intervention operates in terms of cross-attention maps (i.e. attention relationships established between tokens and parts of the image), with the intuition that related tokens should map to the same region of the image, and unrelated tokens to different regions. The paper also introduces a new challenge dataset aimed at diagnosing problems with linguistic binding in image generation (the proposed method performs well on this challenge set).

Strengths

The issue of improper linguistic binding is one that plagues not only attempts at image generation, but many other model uses of grounded language. It's an important problem that has not been adequately resolved by prior work, whereas this paper has demonstrated substantial steps to overcoming this issue without sacrificing on generated image quality. All of this speaks to the significance of this work, which has the potential to see broad and immediate application without the need to undertake costly new model training efforts. The paper is clearly written, and makes effective use of figures to illustrate the problem and how the proposed method resolves it. There is thorough and convincing human evaluation on both existing data, and a new challenge dataset designed specifically to reveal instances of problems with binding.

Weaknesses

There are not many weaknesses to point to. This is not a weakness of the paper per se, but when the method my main remaining (aesthetic) dissatisfaction is the reliance on an external parser, and the possibility of cascading failure that any such pipeline-based system entails.

Questions

none

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Limitations are well addressed in the paper.

Reviewer RH2g7/10 · confidence 4/52023-07-07

Summary

The paper focuses on attribute binding/leak/neglect problems in text-to-image models. The authors propose, SynGen, a method that utilizes dependency trees combined with cross-attention map optimization to achieve better attribute binding results. Extensive experiments are conducted to show the effectiveness of SynGen compared to previous methods.

Strengths

- The paper addresses an important compositional problem in text-to-image generation. - The method is intuitive and effective. SynGen can obtain stronger attribute-object association using dependency trees compared to Structure Diffusion, the positive loss design is well-motivated for solving the problem, and the negative loss design enforces the cross-attention maps over A&E. - The experiment looks comprehensive by considering all sources of datasets and including even more challenging prompts DVMP. The prompts in DVMP are more challenging and realistic than previous "A and B" format prompts.

Weaknesses

- I am concerned with the efficiency of incorporating so many negative losses in the diffusion process. What's the speed of SynGen compared to the original SD? - Using tree-based methods for binding may fail for more complicated and practical prompts. In reality, SD users write much longer prompts that describe the components at different levels. However, I think the contribution of SynGen still matters a lot as the community needs time to develop from methods for short prompts to more generalized methods.

Questions

Apart from the question in weaknesses: 1. Would SynGen still work for prompts like "an apple that is blue" or "a red apple on the left and another on the right"? 2. I may have missed this part but how do you deal with the padding tokens? 3. Table 2 shows that positive only and positive+negative have lower visual appeal percentages than negative only. Does that imply that the positive loss will harm the visual appealing of the images?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See weaknesses.

Reviewer GWNd6/10 · confidence 3/52023-07-09

Summary

This paper approaches the binding problem in text to image diffusion models that is marked by the inability of the models to appropriately identify which modifiers are attached to which nouns in the input text. This often results either in images that mix up the attributes of the different nouns that are mentioned in the text or in images that completely disregard some of the attributes or fall back to statistically likely combinations that are not mentioned in the text. The paper proposes an inference time optimization based on first identifying which modifiers are associated with which nouns, and then optimizing a loss function that explicitly enforces the cross attention matrices of corresponding (i.e. related by modifier-noun relationship) tokens in the text to be similar, while those of every other pair in the sentence to be dissimilar. This results in a diffusion model that adheres more faithfully to the input text and produces images that are deemed better more often than competing baseline approaches that attack the same problem.

Strengths

## Originality, Quality, Significance * The paper provides a novel lightweight method using off the shelf syntactic parsers to enforce better binding of modifiers to nouns in text to image diffusion models. The qualitative results seem promising and the human evaluation results suggest that the proposed method outperforms previous approaches to mitigate the binding issue. ## Clarity The paper is easy to follow. The qualitative examples demonstrate the different kind of issues faced by the baselines and the improvements brought by the proposed approach.

Weaknesses

* Some qualitative examples on the collected DVMP dataset with some categorization on failure cases based on having uncommon modifiers vs different number (performance on 1,2,3 and more modifiers) of modifiers would have been good to see.

Questions

## Clarifications * The metric used in the human evaluation in Table 1 is not sufficiently clear. Aren't the values in each column supposed to sum to 100? * What is the Complex Concepts Prompts dataset mentioned in Fig 5? ## Suggestions * It would be easier for the reader to follow if the examples in Figure 4,5,6 are reorganized such that the references to them in the text are chronological. Currently it requires the reader to keep going back and forth across the figures.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer RH2g2023-08-14

Thank you

Thank the authors for addressing my questions. I would like to maintain my score for now.

Reviewer LWw22023-08-17

Re: Author response

Thanks for the additional experiments and evaluation! There are lots of details brushed under the rug regarding the full evaluation setup but happy to give the benefit of the doubt that these details will be shared and that the evaluation is robust. I've raised my score to reflect this.

Authorsrebuttal2023-08-17

Thank you so much for the support and trust. We will provide all details of the experiments and make our code public to make experiments easy to replicate.

Reviewer GWNd2023-08-19

Thanks for the clarifications

Thanks, and looking forward to the qualitative examples and categorization of errors! Maintaining the score for now.

Reviewer EHJQ2023-08-19

I have read the other reviews, and would like to thank the authors for their responses. With this information in mind, I will maintain my score. Some questions have been raised regarding similarity to the Attend-and-Excite paper. My view on this is that Attend-and-Excite is different in that it doesn't address the attribute binding problem (except indirectly by decreasing attribute neglect). Attribute binding seems to be somewhat resistant to being solved by indirect methods, which contributes to my favorable assessment of the more direct approach taken in this paper.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC