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

Dear reviewer,

We appreciate your valuable suggestions and taking the time to review our work! We have carefully considered your feedback, and we would like to address your concerns. In the following section, we respond to each weakness, while weakness 3 is included in the overall rebuttal at the top of this page. Furthermore, we include corresponding quantitative and qualitative results in Table 1 and Fig. 1 in the rebuttal PDF.

Weakness 1: We acknowledge that the issue occurs less frequently in SD3, as indicated in Table 1 in the rebuttal PDF. While SD3 was announced on Feb. 22, it was only open-sourced on June 12, which was after our paper was submitted. However, the problem is not completely resolved in SD3. In our experiments with the prompt "a lion and an elephant" using 1,000 random seeds, SD3 still exhibited the mixture issue in 2.9% of cases and the missing issue in 3.6% of cases. From SD 1 through SD 3, including SD XL, more text encoders with additional parameters have been integrated into stable diffusion models, with CLIP being a common choice. This underscores the importance of researching how the fundamental CLIP text encoder contributes to T2I models, a topic that has received little attention but is thoroughly researched in our paper.

Weakness 2: Thank you for the suggestion. We analyzed the role of EOT in the Fig. 3 in main paper. In our experiment, when we masked the token embeddings for "a," "cat," "and," "a," and "dog," the remaining embeddings still generated both a dog and a cat in the image due to the information accumulated in the EOT, as discussed in line 154. EOT positively contributes to object coexistence, which aligns with [1] that discusses EOT's role in eliminating negative concepts in prompts. Therefore, we chose not to optimize EOT in our proposed method.

Weakness 4 and 5: We include a quantitative comparison in the rebuttal PDF as Table 1 to compare SD 1.4, SD 1.5, ELLA [3] (on the only released version, SD 1.5), SDXL-Turbo, and SD3 in the color set in T2I-CompBench [4] with 1,000 cases. Within all the baseline methods, our TEBOpt improves the 2 object co-existence rate with 7.9% on SD3 by addressing object mixture and object missing. Also, TEBOpt makes the generated information more balanced, where the info bias is from 2.07 to 1.30. More visual results are provided in Fig. 1 in the rebuttal PDF.

Weakness 6: We carefully consider this feedback and would like to share more thoughts on the decision of the basic model and benchmark.

Initially, we focused on the SD 1.4 model because it contains one commonly-used text encoder, CLIP, the research findings of which could apply to other latest models. For example, current SOTAs, e.g., SDXL and SD3, are equipped with the CLIP text encoder. Regarding the evaluation benchmark, we created our own benchmark for effectively analyzing our proposed problem because there is no suitable benchmark.

To respond, we apply TEBOpt to the latest models, including SD 1.5 with ELLA, SDXL-Turbo, and SD3, on the public T2I-CompBench. We demonstrate that the proposed problem is still not addressed in the latest models, e.g., SD3. Our TEBOpt indeed addresses our proposed problem in all 3 models as well as in the more complex prompt. The visual results are in Fig. 1 in the rebuttal PDF. Specifically, on SD3, our TEBOpt increases 7.9% improvement for 2 object co-existence and makes information bias from 2.07 to 1.30.

Furthermore, our paper conducts a thorough experiment on the contribution of the text encoder to text-to-image models, which is critical but has received little attention. Given the extensive discussion and rapid development of T2I models, and the emerging development of text-to-video models, we believe our task contributes significantly to the community.

References:

[1] Get What You Want, Not What You Don't: Image Content Suppression for Text-to-Image Diffusion Models. ICLR, 2024.

[3] Ella: Equip diffusionmodels with llm for enhanced semantic alignment. ArXiv, 2024.

[4] T2I-CompBench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023.

Dear Reviewer,

We appreciate your valuable suggestions and we are eager to address your concerns with the submitted rebuttal. As the author-reviewer discussion will end soon, we are looking forward to your further consideration of our submitted rebuttal. Thank you again for taking the time to foster a thorough review of our task!

Sincerely, authors

Dear reviewer,

old version SD: The SD3 you suggest is released at 12 Jun., which is later than the date our task submit to NeurIPS. In the rebuttal, we still provide the experiments to prove our effectiveness that our method outperforms SD3. Please refer to the Fig. 1 and Table 1 in the global rebuttal PDF.

latest baselines: In the initial submission, we compared with latest tasks addressing object missing, e.g., A&E (SIGGRAPH’23) [1], SynGen (NeurIPS’23) [13]. To respond to your suggestion, we further compare to Stable Diffusion XL-Turbo, Ella, and SD3. All the qualitative and quantitative results prove our effectiveness. Please refer to Fig. 1 and Table 1 in the global rebuttal PDF.

Benchmark: the public benchmark, e.g., T2I-CompBench, contains more diverse prompts, e.g., adjective binding with noun, that would make the experiment distract. Thus, we follow related works, e.g., A&E (SIGGRAPH’23) [1] and Predicated diffusion (CVPR’24) [16], to create the most suitable benchmark for our task. While we respectfully disagree with the directness to conduct public T2I-CompBench to demonstrate our performance, we still conduct quantitative and qualitative comparisons in T2I-CompBench in the rebuttal and prove our effectiveness. Please refer to Fig. 1 and Table 1 in the global rebuttal PDF.

We sincerely look forward to more concrete details about which part of the rebuttal does not address your concern.

[1] 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. In SIGGRAPH, 2023

[13] 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 NeurIPS, 2023.

[16] Kota Sueyoshi and Takashi Matsubara. Predicated diffusion: Predicate logic-based attention guidance for text-to-image diffusion models. In CVPR, 2024.

Dear reviewer,

We appreciate your valuable feedback and taking the time to review our work! We have carefully considered your suggestions and we would like to address your concerns. In the following section, we respond to each weakness and question, while weaknesses a and c are included in the overall rebuttal at the top of this page. Furthermore, we include corresponding qualitative and quantitative results with more complex prompts for 1,000 samples in Fig. 1 and Table 1 in the rebuttal PDF.

Minor in Strengths: As suggested, we will discuss the related works. Specifically, Zou et al. [A] focuses on image editing based on the cross-attention maps' similarity during denoising process while Pixelface+ [B] focuses on facial manipulation based on text description and segmentation masks.

Weakness b and Question a: Thank you for pointing out the questions about formulas. We clarify every formula related to our analysis (Eq. 3) and proposed method (Eq. 4 to 6) to provide more detailed information.

1. Eq. 3 is the hypothesis equation, expecting to solve the balance problem by replacing the token embedding of the later mentioned object with the corresponding pure token embedding. Due to formatting difficulty in OpenReview, we change Eq. 3 into one line and correct one mistaken condition. \varepsilon' = \varepsilon_i * if * i \notin (O) * or * i = \min (O) * else * \varepsilon_{p_{obj_n},5}, \tag{3} where n refers to the $n^{th}$ object. Eq. 3 shows how the text embedding $\varepsilon'$ combined. Given a prompt "a dog and a cat", there are two critical tokens (<dog> and <cat>) with token indexes 2 and 5 in the critical token set $O$. Since we only replace the later mentioned token with a pure token embedding $\varepsilon_{p_{obj_n},5}$, where 5 indicates the <cat> token index in the pure prompt "a photo of a <cat>", the equation indicates to remain original token embedding when token index i is not in the set $O$ or is equal to the minimum index in the set $O$. In this case, when i is equal to 5 and its corresponding n is equal to 2, the embedding would be replaced.

2. Eq. 4, 5 and 6 are our proposed method's equations: For rendering the correct formula, we omit the "TEB" originally on the right-bottom side for each loss. \mathcal{L}^{pos}(\varepsilon, \varepsilon_p) = \underset{i \in O}{\min} sim(\varepsilon_i, \varepsilon_{p(i)}), \tag{4} \mathcal{L}^{neg}(\varepsilon) = \frac{_1}{^{k(m-1)}} \underset{i \in O}{\sum} \underset{j=1; j \neq i}{\sum^{m-1}} sim(\varepsilon_i, \varepsilon_j), \tag{5} \mathcal{L} = -\mathcal{L}^{pos}+\mathcal{L}^{neg}, \tag{6} where $sim (u, v) = \frac{\mathbf{u}_i \cdot \mathbf{v}_i}{\|\mathbf{u}_i\| \|\mathbf{v}_i\|}$ and m means the effective token count.

• Eq. 4 encourages critical token embeddings $\varepsilon_i$ from given prompt (a dog and a cat) to be more similar to pure critical token embeddings $\varepsilon_{p(i)}$ from pure prompt ("a photo of a <dog>" and "a photo of a <cat>"). We design to select one with minimum similarity to update the text embedding since token with the minimum similarity requires more efforts to purify its embedding.

• Thank you for pointing out the confusion. Eq. 5 encourages critical token embeddings to be less similar to other token embeddings, which has better performance than calculating only with its earlier tokens since the current formula encourages its later tokens not to accumulate all of its information. We believe it is not a conflict with our description since similarity is a mutual result between an earlier token and a later token. Here we calculate the average as the negative loss for update. Also, we correct the equation to "j=1 to m-1" since we do not use <sot> and <eot> for optimization.

• Eq. 6 is the overall loss function, where the lower the better. For achieving this goal, we give a negative sign to the positive loss since it is encouraged to have a larger value/similarity. While we give a positive sign to the negative loss since it is encouraged to have a lower value.

Question b: Thank you for your observation. Compared to the SD 1.4 result in Fig. 4 in the main paper, applying our TEBOpt method significantly reduces the lion’s visual features on the cat, although some features might not be fully eliminated. For example, within the region above the cat's head, distinct and upright strands of hair are eliminated but it still retains some shadow.

The layout of the generated images is controlled by the denoising process, which is influenced by the interaction between the text embedding and the image latent. While we use the same seed for the image latent in all methods with or without our TEBOpt, ensuring the initial image latent remains the same, the optimized text embeddings can still cause changes in the layout. When TEBOpt is combined with A&E or SynGen, it will change the layout compared to SD 1.4 method because A&E and SynGen optimize the image latent during the denoising process. Our primary goal is to address issues of object mixture and omission rather than directly controlling the layout.

References:

[A] 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.

[B] 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.

Dear Reviewer,

We appreciate your valuable suggestions and the positive feedback! We are eager to share further information about our task with you in the submitted rebuttal. As the author-reviewer discussion will end soon, we are looking forward to your further consideration of our submitted rebuttal. Thank you again for taking the time to foster a thorough review of our task!

Sincerely, authors

Dear reviewer,

We sincerely appreciate your valuable suggestions and supportive feedback, as well as the time you took to review our work! We love your summary of our work, demonstrating a thorough understanding. In the following section, we respond to each weakness and question, while weakness 3 is included in the overall rebuttal at the top of this page.

Weakness 1: Thank you for the suggestion. We will enhance the aesthetics of the table.

Weakness 2: Thank you for the suggestion. We have already built an automatic data generator and an evaluation metric, and we plan to extend and organize the data for community use. In the supplementary materials, you can find the code for systematically extending the data in 'code/data/gen_prompt.py', and the code for the automatic evaluation metric in 'code/eval_metrics'.

Question 1: Thank you for your insightful comment. We explored this question when we hypothesized the solution to our proposed problem. We conducted an experiment on the same 400-prompt set described in the paper, eliminating "all" earlier information from accumulating in subsequent tokens. The results are presented in the table below. Without shared embeddings across tokens, the generation process failed to produce co-existing objects. Specifically, during the denoising process, each object token responded in the central region, as observed in the cross-attention maps, resulting in no object co-existence. In conclusion, maintaining a proper proportion of earlier object token information in the latter tokens (excluding those with concrete meanings) has more positive than negative effects, especially in generating co-existing objects within a given prompt. Therefore, we propose to optimize the critical tokens' embeddings in the paper.

Dear Reviewer,

We appreciate your recognition of our work as a valuable starting point! Thank you also for your thorough and thoughtful comments. We will incorporate your valuable suggestions into an updated version and continue contributing to the community.

Thank you again for your support!

I appreciate the authors' efforts in addressing my concerns. After reviewing the authors' rebuttal and the comments from other reviewers, I maintain my initial stance that this paper serves as a valuable starting point in this direction and offers insights into the phenomenon through error analysis and initial results. Therefore, I will keep my original rating as 'accept.'

However, there is still much room for improvement. As noted by all the reviewers, the current version examines the problem in a relatively simple setting, and there are opportunities to further explore the problem setting, model choices, and evaluation benchmarks more systematically (though some have been provided in the rebuttal). I hope the authors can continue refining their work to better benefit the community.

Dear Reviewer,

Thank you for taking the time to review our work and providing us with your valuable suggestions and insightful comments. We have carefully considered your feedback, and we would like to address your concerns. In the following section, we respond to each weakness (W) and question (Q), while weakness 1 is included in the overall rebuttal at the top of this page.

W2: We agree that 2 objects in the real-world overlapping >90% might not mean mixture, such as overlapping persons. However, in diffusion-based generated images, this high overlap does indicate mixtures due to the nature of the models used. Considering the working mechanism of SOTA text-to-image models, when the cross-attention maps of 2 objects respond closely during denoising, there is a high probability of generating a mixture object, as Fig. 2 in the rebuttal PDF shown. Consequently, within the nature of the Owl-ViT detector, these mixtures can be identified by their high overlap.

W3:

• Clarification for Information bias and object missing: Due to the causal masking manner, the earlier token information accumulates to all the later token embeddings, making the overall text embedding biased towards the token information mentioned earlier in the text prompt. We call this scenario information bias. Then, biased text embedding, having more information of the earlier mentioned token, makes the denoising process to have higher probability to generate images with the object mentioned earlier and leads to object missing.
• Thank you for raising the question. We will change the term critical "dimension" to critical "token embedding". Take the prompt “a dog and a cat” as an example. There are 2 critical tokens, <dog> and <cat>, and we generate the pure text embeddings for them by the prompts "a photo of a <dog>" and "a photo of a <cat>", respectively. Then, we use the <dog>'s and <cat>'s token embeddings as critical (pure) token embeddings for calculating positive loss in TEBOpt.

W4: Our TEBOpt optimizes the text embedding based on pure token embedding and original token embedding "before" the denoising process while the text embedding is fixed during denoising process for fixing the generated meaning guidance. Regarding the “initial” image latent mentioned in L#197, we will polish the wording to reduce confusion.

W5: The baseline in Fig. 5 is SD 1.4.

W6: No problem. Please refer to Fig. 1 in the rebuttal PDF.

W7: We statistically calculated the p-value by the chi-square test in Table 4 to prove significance. The p-values for SD 1.4 vs. SD 1.4 + TEBOpt on info bias and overall performance, including 2 objects exist, object mixture, and object missing, are 0.0052 and 0.0108, which are far below the alpha value of 0.05, proving significant improvement.

W8: As suggested, we will discuss the related works. Specifically, [composable] separates the given prompt into different objects and generates images using a set of diffusion models. [custom diffusion] focuses on preserving the customized concept for text-to-image (T2I) generation while its limitation (Fig. 9) is exactly our main focus that pretrained T2I models struggle with similar compositions, e.g., dog and cat. [Hiper] reconstructs text embedding for image editing, while its main target for image editing is different from our task.

W9: Compared to retraining the text encoder directly, our approach is highly cost-efficient as it employs a training-free inference design, serving as a proof of concept. Specifically, our optimization adds only 3.11% to the inference time, based on an average of 1,000 cases using a single A100 GPU.

Mics (M) 1: Thank you for the suggestion. We will make the tables consistent.

M2: To improve numerical readability, we acknowledge that 125.42% is calculated in 2 steps. The info bias value indicates less bias when it is closer to 1, different from the intuitive assumption that lower values represent less bias. Here’s the detailed calculation:

1. We calculate the bias distance between the info bias value and the balanced value (which is 1). The bias distances of SD and SD + TEBOpt are (2.647-1)/1 = 164.71% and (1.393-1)/1 = 39.29%. Note that we have rounded the values for information bias in Table 4, reporting 2.647 as 2.65 and 1.393 as 1.39.

2. The balance improvement: 164.71%-39.29% = 125.42%.

M3: The balance performance is assessed by how close the information bias is to 1, as this indicates a more balanced state. Thus, SynGen has a lower info bias value while SynGen + TEBOpt is indeed improved from 0.62 to 0.72 because 0.72 is closer to 1.

M4: We considered that full prompt similarity is insensitive to numerical information. To describe the case more properly, we will update the image with 1 cat as in Fig. 3 in the rebuttal PDF. Still, it has a lower score than the mixture one in Fig. 7, indicating the inadequacy of the metric.

Q1: We experimented with 1,000 samples on the effect of words that may change their meaning due to nearby words, including "mouse", "horn", "jaguar", "falcon", and "palm". Specifically, we use the prompt "an <animal/object A> and a <B>" and evaluate the result by detecting 2 targets <animal A> and <object A> using Owl-ViT detector. Our TEBOpt can address 3.67% object missing issue in animal prompts while the optimized results may lean towards the main meaning of the word in object prompts. For example, "jaguar" tends to represent an animal rather than a car, resulting in a 1.29% decrease in generating "object jaguar" after optimization.

Q2: While we agree that "dependency" is a more accurate term, the term "causal" in this context refers to the causal masking manner used in the self-attention layer. This is described on P. 5 of the CLIP paper, where the source code function is also named "causal_attention_mask" (line 288 in modeling_clip.py). Thus, we used "causal" to maintain consistency with the terminology used in the CLIP documentation.

Dear Reviewer,

We appreciate your valuable suggestions and we are eager to address your concerns with the submitted rebuttal. As the author-reviewer discussion will end soon, we are looking forward to your further consideration of our submitted rebuttal. Thank you again for taking the time to foster a thorough review of our task!

Sincerely, authors

Dear our reviewers,

Now, we are approaching the very end of the reviewer-author discussion period. Are there any further questions for the authors regarding major issues, or are your concerns sufficiently resolved? Only a limited amount of time is available for interaction with the authors.

Thank you again for your dedication.

Your AC