Exposing Text-Image Inconsistency Using Diffusion Models

Q1. Could this method be extended to evaluate other text-to-image (T2I) augmentation techniques?

Thanks for the suggestion. Running these experiments takes time so we cannot do it within the period of ICLR review. Technically, our method is not restricted to a specific text-to-image diffusion model, and can be extended to other most recent diffusion models and their variants.

Q2. Are there alternative evaluation metrics to assess the correspondence between text and images other than CLIP scores?

Following existing text-image correspondence evaluation methods [1-4], we used the CLIP score to assess the correspondence between text and images. It is worth noting that the CLIP score is only used to analyze the inconsistency levels of our dataset in Table 1 and evaluate the text embedding initialization effectiveness in Table 7, it might work well on our dataset as most of the inconsistencies in our dataset are in the object-level but not scenarios-level. We did not use the CLIP score in the text-image inconsistency localization pipeline, which will not influence the localization performance.

[1] Meng, Chenlin, et al. "On distillation of guided diffusion models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

[2] Saharia, Chitwan, et al. "Photorealistic text-to-image diffusion models with deep language understanding." Advances in Neural Information Processing Systems 35 (2022): 36479-36494.

[3] Couairon, Guillaume, et al. "Diffedit: Diffusion-based semantic image editing with mask guidance." ICLR. 2023.

[4] Blattmann, Andreas, et al. "Retrieval-augmented diffusion models." Advances in Neural Information Processing Systems 35 (2022): 15309-15324.

Q1. The general theoretical idea behind the approach lacks clearity.

Thank you for your suggestion on the paper. We will improve the writing in the methodology part to make it more readable. The theoretical idea behind the approach is to employ the text-to-image diffusion models as “omniscient” agents to align the image-text latent space so that our proposed two alignment steps are able to filter out irrelevant information and incorporate background knowledge to identify inconsistencies.

Q2. The real-world application is not very clear.

We describe the problem that this paper solves in the Introduction. Our method aims to expose misinformation on social media and the Internet created by juxtaposing images with texts that do not accurately reflect the image’s original meaning or intention.

Q3. Sensitivity to mask threshold.

We agree that the mask threshold influences the performance. Therefore, our method designs a sample-specific threshold based on the average values among the mask instead of using a fixed threshold for all samples. In Table 10 and Figure 12 of the Appendix, we conducted a comparison with four fixed threshold strategies to show the effectiveness of our mask threshold setting method.

Q4. Deeper insights into why this approach works are lacking.

Understanding the semantic connection between text and image is the key to solving the text-image inconsistency problem, and our method is the first to explore generative AI models for this purpose. Our results have revealed two aspects that are not known previously in the literature. First, we use the large-scale text-to-image diffusion models as a foundation model with extensive background knowledge to effectively align the text and image semantics. Second, we design two alignment steps to iteratively align the image/text (embeddings) and filter out relevant semantic information with diffusion models. Thus, we obtain knowledge-shared text embeddings from both the input image and text, making it easier to identify semantic inconsistencies.

Q5. Lack investigate the method in more detail in terms of learned representations.

We have included the investigation of the learned representations in the revised manuscript of paragraph “Analysis of the learned representation” in the Appendix.

Q6. How the approach works on other datasets,is there another dataset where the method could be compared?

To the best of our knowledge, the proposed TIIL dataset is the first text-image inconsistency localization dataset with pixel-level and word-level inconsistency annotations. The proposed TIIL dataset is based on content from diverse real-world news stories and a wide range of modifications in inconsistent regions, including local inconsistencies and global swapping based on mixed-up pairs.

Q7. Could the model be sensitive to the DALLE generated part instead being sensitive to the text-image inconsistency?

Our method is not sensitive to the part that DALL-E generated. As shown in the Appendix, we compared the performance of our method with baselines on Image-manipulated and Text-manipulated subsets in Tables 8 and 9. The results show the effectiveness and superiority of our method in both DALL-E generated inconsistent samples and non-DALL-E manipulated samples.

Q8 What are the shortcomings of proposed methods?

As discussed in Section 5.4 on failure cases, one shortcoming of our method is that given the limited prior knowledge of the diffusion model we used, our model may not effectively handle the inconsistencies with respect to specific external knowledge.

Q9. How does the approach perform on completely wrong image descriptions?

We have added examples with completely wrong image descriptions from our TIIL dataset in the Appendix Fig.13 (a) of the revised manuscript. Our method can provide a whole image mask for this kind of inconsistency samples.

Q1. D-TIIL may struggle with inconsistencies with respect to specific external knowledge which could reduce the effectiveness of the method in real-world application.

With the rapid development of text-to-image (T2I) diffusion models, our D-TIIL would be more generalizable to handle different types of real-world cases, such as using domain-specific diffusion models as mentioned in the Conclusion or more powerful diffusion models like the Stable Diffusion XL.

Q2. The dependence with diffusion model could limit the generalizability of the proposed method, especially when text and image are not well aligned on the latent space.

Our method leverages the semantic alignment performance of T2I diffusion models, but would fail for complex cases that current diffusion models cannot handle. However, our method is not restricted to a specific T2I diffusion model and can be implemented and adapted to the most recent diffusion models to enhance its generalizability further. In terms of the generalizability to the capacity for identifying various degrees of inconsistency in text-image pairs, our current method can address not only subtle inconsistencies with local manipulations but also whole swapping and mix-ups. We provided some examples of such scenarios in Fig. 13 (a) and (c) of the revised manuscript.

Q3. There are some confusing details in the method description section.

Thanks for pointing this out. We will improve the writing of the methodology part in the finalized version of the paper.

Q4. The reasons why D-TIIL is superior are not discussed and analyzed in detail, and the potential solutions for the failure cases are not provided.

Our D-TIIL outperforms baselines in comparison experiments due to the following two reasons. Unlike the baseline DetCLIP using object segmentation to compare CLIP similarity or GAE using attention heatmap for CLIP, our method leverages the T2I diffusion models to learn the semantic connections between textual and visual information. More importantly, instead of directly comparing the text and image embeddings, D-TIIL relies on two-step alignment to iteratively exclude irrelevant information and obtain knowledge-shared representations from two modalities and then directly exposes the inconsistencies. The potential solutions for failure cases involve using more powerful diffusion models such as the most recent T2I diffusion models or domain-specific diffusion models, as discussed in the Conclusion section of the submission.

Q5. More specific discussions and measures could be included to prevent potential abuse rather than simply restricting access.

We will release our code as open-source with the condition that it “must not distribute harmful, offensive, dehumanizing content or otherwise harmful representations of people or their environments, cultures, religions, etc. produced with the model weights”.

Q6. The term “text denoising” is too idealistic and D-TIIL may not localize the inconsistency that backgrounds or objects that contain more complex semantics.

Thanks for the suggestion, we will consider using the term “text alignment” other than “text denoising”. This process actually reduces the distance between the image and text semantic space as shown in the "Analysis of the learned representation" section of the Appendix in the revised manuscript. Our method can handle inconsistencies in not only simple objects, but also the backgrounds, and attributes of objects. We have included such examples in Fig. 13 (c) of the Appendix.

Q7. The descriptions of Step 4 in Method is not clear.

Thanks for pointing this out. We have modified the statement as follows: “We then compute the cosine similarity score between the CLIP image embedding of the masked image and the input text embedding E_0 as the consistency score.”.

Q8. IN Section 4 DATASET, T_{m} is unknown.

T_{m} is the alerted text based on human annotations. We have clarified this in Section 4. Thanks for pointing this out.

Q1. The writing needs to be improved.

Thank you for your suggestion on the paper. We will improve the writing in the finalized version of this manuscript and pay special attention to the methodology part to make it more readable.

Q2. How does the annotation and the model handle predicates inconsistencies?

This is a good point. Our method can well handle inconsistencies in not only objects but also predicates and adjectives. We have added examples in the Appendix Fig. 13 of the revised manuscript. In terms of dataset construction, the annotators are instructed to select and edit object-term pairs from image and text, which include editing in objects, scenes, and attributes based on predicates or adjectives. For example, a singing man is edited with the prompt “a man playing basketball”, and a yellow cat can be changed with the prompt “a red cat”.