A Cat Is A Cat (Not A Dog!): Unraveling Information Mix-ups in Text-to-Image Encoders through Causal Analysis and Embedding Optimization

Summary

This papers address the bia problem of text-to-image generation. Unlike the previous work which focus on processing of the attention map, this work explores the text embedding, which is a key factor. Authors conduct expensive experiments to analysis the text embedding, then propose a interesting method to address the alignment of the text-to-image generation. A simple text embedding alignment loss is proposed by introducing other pure text information.

Strengths

1. Authors conduct lots of interesting experiments to analysis the alignment problem, including masking the special token. 2. Exploring the text embedding instead of (stable diffusion ) UNet network is good. 3. Authors provide the code to show the effectiveness of the presented method.

Weaknesses

1. The present problem (e.g., Fig1) seems just happy the lower version of SD1. I test SD3 with the prompt : 'a lion and an elephant', the issues does not happen. In the paper, authors show the reason is from the text encoder, which I agree. However, the presented reason maybe does not lead to the shown challenges. 2. For text embedding, the EOT also play a key role[1]. This paper does not consider it. Only considering the soft replacing maybe harder. 3. The visual results are not good. The generated images has poor quality. Does the propose lead to other negative influence? 4. The baselines are a little out of fashion. Authors should consider more recent methods: SDXL, SD3, PlayG-v2, Ranni[2] and Ella[3]. 5. Authors should use a public benchmarks to compare the proposed method, like T2I-CompBench[4]. 6. The proposed idea seems limited for NeurIPS, which is not enough. [1] Get What You Want, Not What You Don't: Image Content Suppression for Text-to-Image Diffusion Models [2] Ranni: Taming text-to-image diffusion for accurate instruction following. [3] Ella: Equip diffusionmodels with llm for enhanced semantic alignment. [4] T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation

Questions

Please check the weaknesses.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

Testing the presented problem in latest version of SD, like SD3.

Summary

The paper conducts an in-depth analysis of the impact of causal mechanisms in text encoders of text-to-image (T2I) diffusion models, which can lead to information bias and loss. The authors propose the Text Embedding Balance Optimization (TEBOpt) method, a training-free method to optimize text embeddings, resulting in a 90.05% improvement in information balance. Additionally, it presents a new metric for more accurate evaluation of information loss in generated images.

Strengths

a. Originality: The paper offers a new perspective on T2I models by addressing the causal nature of text embeddings, which has been overlooked. The proposed TEBOpt represents a new approach to balancing information in text embeddings without additional training. b. Quality: Technically sound, the paper is well-supported by comprehensive analysis and experimental results. c. Clarity: The submission is clearly written and well-organized, making complex concepts accessible. d. Significance: The results are important for advancing the field of T2I models, offering a demonstrable improvement in handling information bias and loss. minors: some relevant works should be cited and dicsussed: Zou, Siyu, et al. "Towards Efficient Diffusion-Based Image Editing with Instant Attention Masks." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 7. 2024. Du, Xiaoxiong, et al. "Pixelface+: Towards controllable face generation and manipulation with text descriptions and segmentation masks." Proceedings of the 31st acm international conference on multimedia. 2023.

Weaknesses

a. The paper has limited qualitative results and mentions that the analysis was conducted on a total of 400 prompts. Providing more qualitative results in the main text or appendix could better demonstrate the validity of the proposed method. b. Some of the formulas used in the paper are not thoroughly explained, making it difficult to understand the proposed method. c. The paper proposes a novel perspective to analyze object omissions and fusion issues in the T2I model. However, the method in the paper is rather straightforward, which to some extent may affect the quality of the generated images.

Questions

a. According to the description in the text, Equation 6 is intended to encourage later tokens to be as dissimilar as possible from earlier tokens’ embedding. However, in the formula, the similarity is calculated between the object token and all tokens, with the summation range being 0 to n. This seems to contradict the goal of constraining the subsequent tokens to be dissimilar from earlier tokens? b. In the qualitative results shown in the second row of Figure 4, it can be observed that the results using the TEBOpt method do not seem to retain the visual features of the cat very well. Does this indicate that optimizing and adjusting text embeddings is likely to cause some uncontrollable changes in semantics? Additionally, when combined with other methods, it also seems to affect the overall layout of the generated images.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper summarizes the limitations of the proposed method in sec6, while mentioning that the case of multiple objects as well as complex prompts can be further explored. In addition, the experiments in this paper are based on simple sentences, which can be followed up to explore whether the case with complex adjectives affects the generation effects.

Summary

This paper proposes Text Embedding Balance Optimization (TEBOpt) to enhance the distinctiveness between text embeddings of equally important objects when using texts as conditions for diffusion models. It begins with an intriguing and informative experiment to investigate why a prompt like "An <object1> and an <object2>" given to a standard diffusion model, such as Stable Diffusion, often results in a higher likelihood of ignoring <object2> during image generation. The masking experiment elucidates the role of causal masking in contributing to this phenomenon. Based on these observations and analyses, the authors propose replacing the latter embedding with a pure embedding as a baseline. They then learn to increase the distance between the embeddings of the two objects while preserving the original meaning. Both qualitative and quantitative experimental results demonstrate that the proposed method helps to reduce bias towards the first object, leading to improved generation results, even compared to existing approaches that aim to address this issue.

Strengths

- The method is well-motivated by the detailed error analysis and masking experiments. - The solution proposed by the authors are novel and interesting. - Experimental results show that the method successfully help to improve the generation quality. - The paper is well-organized and easy-to-follow. - Code and models are provided to better support the reproductivity.

Weaknesses

- Some of the wrapped table's layout can be improved (e.g. Table 1, Table 5) - The data used in this work can be systematically extended and organized to form a larger and good place for community to continue conducting research on this. - The research by far focuses on a few relatively simple scenarios, where only two or three objects in a very simple sentence pattern are presented. Explorations on more challenging scenarios would be interesting.

Questions

- As revealed by the analysis, not only the latter object text embedding contains the earlier object information, all the latter words all may have similar impacts to strength the text embedding of the earlier object information. Has the author considered how to resolve such possible influence?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are potential impact are discussed in the paper.

Summary

Background: Text-to-image diffusion models, such as Stable Diffusion, frequently encounter difficulties in accurately generating images from textual descriptions, especially when there are multiple objects -> object mixing and missing. This paper focuses text embeddings to solve the problems. Hypothesis: self-attention mixes text embeddings of multiple objects. Analyses: * 400 prompts of "a/an <animal 1/2> and a/an <animal 2/1>" with 17 animals produce lots of mixture of objects and missing objects * Contribution of text embeddings to the generated images * Masking the dominant embeddings reduces the mixture rate * Replacing the embedding after the text encoder with the pure text embedding does not solve the problem. * Process of conditions getting lost and biasing towards the first-mentioned object due to masks in self-attention in the text encoder. Method: * Optimize the target text embedding toward pure embedding and against earlier (first-mentioned) token’s embedding

Strengths

This paper provides useful analyses on text embeddings for text-to-image diffusion models. The experiment design corresponds to the hypothesis. The proposed metric is a good effort for quantitative evaluation.

Weaknesses

(Overall, I like the idea of this paper but it needs more work to be a solid paper.) W1. This paper covers only nouns, especially animals. The research community has tackled different aspects of text prompts [composable, masactrl]. How about using other classes of objects to prove generality? For instance, could this method work with text prompts like 'A <cat> standing next to the <tower>' or examples from A&E such as 'A <lion> with a <crown>'? It would be valuable to see if the approach is equally effective with these more diverse classes and prompts that users might want to apply. W2. Bboxes overlapping more than 90% not always indicate mixture. W3. The terms should be more clearly defined. * Information bias and object missing are confusing. * The critical dimension? W4. Initial latent optimization is poorly described. How can the initial latent be updated during the denoising process? I guess intermediate latents can be updated just like [self-guidance]. W5. The baseline of Figure 5 is not defined. W6. The appendix should provide large number of examples to ensure the proposed method is free from cherry picking. W7. Improvements in Table 4 is marginal. W8. An important paper is missing: [composable], [custom diffusion] and [hiper] W9. The method requires optimization for each sentence, which can be inefficient for practical use. Misc. 1. Placing the caption of Table 5 under the table would be better for consistency. 2. Some numerical values in the main text are difficult to understand at first glance. To improve readability, I think it would be better to highlight or mark the key values within the table itself. For example, the main text mentions 125.42%, but I couldn't immediately identify which part in Table 4 shows a difference of 125.42%.  3. The result shows that the balance performance in SynGen decreased by 10.65%, rather than improved. 4. I think Figure 7 is not an appropriate example for demonstrating the inadequacy of full prompt similarity, as neither image correctly matches the given prompt 'a cat and a penguin'. The left image shows a mixture of the two animals, while the right image displays two cats instead of one. [composable] Compositional visual generation with composable diffusion models, Liu+ [masactrl] MasaCtrl: Tuning-Free Mutual Self-Attention Control for Consistent Image Synthesis and Editing, Cao+ [self-guidance] Diffusion Self-Guidance for Controllable Image Generation, Epstein+ [custom diffusion] Multi-Concept Customization of Text-to-Image Diffusion, Kumari+ [hiper] Highly Personalized Text Embedding for Image Manipulation by Stable Diffusion, Han+

Questions

Q1. Would the same happen for the words may change their meaning due to nearby words? E.g., mouse? Q2. I do not think the self-attention layer models "causal" relationship between words. I recommend replacing "causal manner" with "dependency". Thoughts?

Rating

6

Confidence

5

Soundness

2

Presentation

2

Contribution

3

Limitations

Some are covered.