Image Copy Detection for Diffusion Models

Images produced by diffusion models are increasingly popular in digital artwork and visual marketing. However, such generated images might replicate content from existing ones and pose the challenge of content originality. Existing Image Copy Detection (ICD) models, though accurate in detecting hand-crafted replicas, overlook the challenge from diffusion models. This motivates us to introduce ICDiff, the first ICD specialized for diffusion models. To this end, we construct a Diffusion-Replication (D-Rep) dataset and correspondingly propose a novel deep embedding method. D-Rep uses a state-of-the-art diffusion model (Stable Diffusion V1.5) to generate 40, 000 image-replica pairs, which are manually annotated into 6 replication levels ranging from 0 (no replication) to 5 (total replication). Our method, PDF-Embedding, transforms the replication level of each image-replica pair into a probability density function (PDF) as the supervision signal. The intuition is that the probability of neighboring replication levels should be continuous and smooth. Experimental results show that PDF-Embedding surpasses protocol-driven methods and non-PDF choices on the D-Rep test set. Moreover, by utilizing PDF-Embedding, we find that the replication ratios of well-known diffusion models against an open-source gallery range from 10% to 20%. The project is publicly available at https://icdiff.github.io/.

Paper

References (52)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer vdEZ6/10 · confidence 4/52024-07-07

Summary

This paper studies a novel task of Image Copy Detection for Diffusion Models (ICDiff). Different from traditional Image Copy Detection task, ICDiff focuses on detect and evaluate the degree of image copy for images generated by text-2-image (T2I) generative diffusion models. This is an important and meaningful task for the community to study, with the emergence and influence of T2I diffusion models on downstream tasks, such as image editing, icon generation, or style transfer. Effectively and precisely detecting the generated images that might "copy" content/style/character of the artifically created images in its training data can help the end users avoid violating issues such as Image Copyright. To study such a problem, the authors collected and proposed a new dataset called D-Rep. To build such a set, the authors first try to find text prompts from DiffusionDB that are highly similar to image titles from LAION-Aesthetics V2, using Sentence-Bert score. Subsequently, they generate 40k images using these collected prompts using Stable Diffusion 1.5 to simulate possible image copies from LAION. To obtain the degree of image copy annotation, the authors employ human annotators to score each generated-image-&-LAION-image paris from 1 to 6. Furthermore, the authors propose and train a probablity density based embedding model (PDF-Embedding) to score the degreee of image copy for input images. 90% of the collected D-Rep dataset is used as training set for the proposed method. In the experiments, the proposed PDF-Embedding more effectively and efficiently detect image coplies in the D-Rep test set, comparing to the compared methods. However, the quantative experiments are limited to the proposed dataset. In addition, the comparison is somewhat unfair because many of the compared methods (such as GPT-4V, CLIP, DINOv2) are not trained or fine-tuned on the training split of D-Rep, but the proposed method is.

Strengths

- [Clear writing]: The writing is clear. The sturcture of the paper is easy to follow. - [Novel and Important Task]: The studied and proposed task Image Copy Detection for Diffusion Models (ICDiff) is novel, practical, and meaningful. It is important for the community to address the Image Copy issues caused by the training data of T2I diffusion models. This will significantly help downstream tasks avoid copyright issue. - [Good baseline]: The proposed PDF-embedding is an efficient and good baseline to address the proposed task.

Weaknesses

- [Incompleteness]: This paper is not complete in itself. The authors largely use "refer to Section XXX in the Appendix" throughout the paper. At Sec. 5.1 (Line#185), the entire section for Training Details contains only 1 sentence "Please refer to Section E in the Appendix". I understand the space is limited in the main manuscript and the authors want to use the space for more important content. However, a paper should be complete in itself. After carefully reading the paper, I cannot know: i) how the model is trained, what is the architecture, what are the configuration; ii) how the other methods are compared, do the authors train them, how are they trained or adapted to the current task? These details are important for the main paper for the readers to understand the method and the experiments. - [Limited experiments]: The quantitaitve experiments are only conducted on one dataset (4k images in the test set) in Tab. 1. In addition, this dataset is generated by only one specific T2I model Stable Diffusion 1.5. Different generative models might have different image copy patterns. The current experiments can hardly validate the effectiveness and advantages of the proposed method over other methods. More comprehensive experiments should be conducted. When switching prompts (data) and the generative models, maybe zero-shot methods such as CLIP, M-LLMs (GPT4-V), or DINOv2 are more generalize. Without large scale experiments, it is hard to draw a conclusion. - [Unfair comparison]: In Tab. 1, the compared Vision-Language Models (CLIP, GPT-4) and Self-supervised Learning Models (e.g., DINOv2) are not trained or fine-tuned on the evaluation dataset's training split. However, the proposed method is trained on the training split. This makes the comparison unfair. To ensure a fair comparison, either the compared methods should be trained or fine-tuned on the same training split; or, evaluate all methods on novel datasets in a zero-shot manner. With the current unfair comparison, the advantage and effectiveness of the proposed method is not clear. - [Duplicate content]: Sec. 5.2 and 5.3 are highly overlapped.

Questions

Q1: Do the authors study the quality of text matching using Sentence-Bert? Pure linguistic matching could be quite ambigious and subseqeuntly leads to incorrect matching. For instance, "containers" can be matched to both the huge containers for shipment at the port and the plastic containers (such as lunch box) in a kitchen context. In addition, do the authors study different scores, such as CLIP text encoder score instead of Sentence-Bert score? As observed in [1], for visual content related text, CLIP encoder seems better than Sentence-Bert. Q2: How do you define Image Copy for Diffusion Models? What kind of "Copy" should be detected? This is important for the studied task. For instance, when a user prompts a T2I diffusion model "A portrait of the Mona Lisa.", of course, a highly similar image that copies the famous Mona Lisa potrait will be generated. But, is this an image copy? This kind of detection might not helpful for downstream tasks. The task should be defined better and thereby the method can be designed better. From the visualization in Fig. 3, I found the high-score copies are mainly "object copy". These copies might occur frequently if the prompt tasks the diffusion model to generate such an image. By conditioning and limiting the prompt text, this can be solved directly. So why is it useful? T In my opinion, a worth-to-detect image copy would be when prompting "Generate a yellow cartoon-style mouse" and a Pikachu-like image is automatically generated without explicitly mentioning it in the prompt. Q3: How well the proposed method will generalize to images generated by other diffusion models, such as SDXL or DALL-E-3, quantatively? [1] Chen, Z., Chen, G. H., Diao, S., Wan, X., & Wang, B. (2023). On the Difference of BERT-style and CLIP-style Text Encoders. arXiv preprint arXiv:2306.03678.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer 2jDa7/10 · confidence 5/52024-07-10

Summary

This paper constructs a Diffusion-Replication dataset aiming to solve the image copy detection problem for diffusion models. This paper proposed a strong baseline named PDF-Embedding which transforms the replication level into a probability density function as the supervision signal. Extensive experimental results and analysis demonstrate the efficiency of the proposed method.

Strengths

1. The topic is timely and interesting. The presentation is clear and easy to follow. 2. This paper creates a valuable dataset, which is important to identify the replication caused by diffusion models. 3. This paper gives a reasonable theoretical explanation for the proposed method. The analysis of the proposed PDFs is convincing on three primary functions: Gaussian, linear, and exponential. 4. The experiments are intensive and reasonable results are achieved.

Weaknesses

1. As shown in Figure 5 and Figure 15, authors use different values of A for each function, while there is no explanation on both the selection rule and influence of A. 2. As shown in Table 2, the training time, inference time and matching time are not better than other methods. 3. This paper creates a valuable dataset, while the details of the data set, such as the label distribution of the dataset, are missing, which is important in practical applications.

Questions

1. Authors use different values of A for eq3-5, but there is no explanation on both the selection rule and influence of A. 2. The details of the data set, such as the label distribution of the dataset, are missing.

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes, the authors addressed the limitations and potential negative societal impact of their work.

Reviewer dJ327/10 · confidence 3/52024-07-12

Summary

This paper introduces a novel method named ICDiff for detecting whether images generated by diffusion models replicate the training set. The authors have constructed a new dataset called D-Rep and proposed a new embedding method called PDF-Embedding. This approach transforms the replication level of image pairs into a probability density function, using this as a supervisory signal to train the model. The experimental results demonstrate that ICDiff outperforms many existing image copy detection method.

Strengths

1. Innovation: ICDiff is the first image copy detection method specifically aimed at replicas generated by diffusion models, filling a gap in current research. 2. Dataset Construction: The creation of the D-Rep dataset provides a valuable resource for the study of image copy detection, with its replication level annotations offering clear guidance for model training and evaluation. 3. Methodology: The PDF-Embedding method, which utilizes a probability density function as the supervisory signal, is both innovative and effective.

Weaknesses

See questions.

Questions

1. Intuitively, we would use a single set of vectors to characterize the similarity between images, with larger dot product values indicating higher similarity and smaller ones indicating the opposite. This paper, however, uses six sets of vectors to measure six levels of similarity, which seems counterintuitive. Specifically, when two images are almost identical, the set of vectors indicating their similarity as zero needs to be quite different; when two images are completely unrelated, the set of vectors indicating their similarity as zero needs to be as consistent as possible. I am curious about the rationale behind this design. 2. Although ICDiff performs well on the D-Rep dataset, its generalizability to other datasets or different types of diffusion models has not been verified. In particular, I am concerned whether other methods compared with ICDiff (such as SSCD) have been trained on the D-Rep training set? If so, how was the training conducted; if not, is this comparison on the D-Rep test set somewhat unfair? 3. Regarding the design of the Relative Deviation (RD). The authors point out in Appendix B that when $s^l=3,s^p=0$, the penalty should be greater than when $s^l=5,s^p=2$; my question is, when $s^l=5,s^p=0$, should the penalty be greater than when $s^l=3,s^p=0$ (although both can't be more wrong, the former seems to be more egregiously wrong)? 4. How do the authors utilize the trained ICDiff to assess the replication ratio of diffusion models? The paper seems to only mention conclusions (such as 10.91% or 20.21%) without explaining how these figures are derived. Specifically, ICDiff provides six replication levels for each set of images. When assessing the replication ratio of diffusion models, what criteria or which level does the author use as a threshold?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

No limitations.

Reviewer dJ322024-08-08

Thank you for your detailed response. You summarized my questions as Q1-Q5, which I find to be appropriate. Regarding Q2, Q3, and Q5, I have no additional queries. For Q1, I would like to delve deeper into a hypothetical scenario. Imagine figures A and B are identical (or almost identical), and we have a vector derived from A and another from B following identical procedures. In such a case, these vectors should have a minimal dot product. How does the model address this requirement? For Q4, I would appreciate further clarification on the statement, "we cannot normalize the penalty to the range $0 \sim 1$ for each pair, which may lead to overly optimistic results." Why will we have overly optimistic results? Thank you again and look forward to your further response.

Authorsrebuttal2024-08-09

Thanks and further clarification

Thank you for your prompt response and for participating in the rebuttal discussion. We are happy to have addressed 3 out of 5 questions and would like to further discuss your follow-up concerns. **Q1 Further. Imagine figures A and B are identical (or almost identical), and we have a vector derived from A and another from B following identical procedures. In such a case, these vectors should have a minimal dot product. How does the model address this requirement?** A1 Further. Thank you for this insightful question. If images A and B are almost identical, their vectors at Level 0 are indeed very similar, resulting in a high inner product. However, the inner product at Level 5 is even larger. Therefore, because our method relies on the relative scale relationship — using the highest-scoring entry to indicate the predicted replication level — it remains effective. For instance, the second subfigure in the first row of Fig. 16 shows two almost identical images, where the original dot products for Levels 0 and 5 are 0.864 and 0.998, respectively. Our method successfully predicts this pair as Level 5. **Q4 Further. Why will we have overly optimistic results?** A4 Further. Sorry for the confusion, and we clarify this as follows: Our statement was to highlight that the Relative Deviation is more intuitive because it normalizes the deviation: assigning 1 deviation score to the worst case and 0 deviation score to the best case. In contrast, Absolute Deviation does not provide this intuitive scale. For example, Absolute Deviation for the case where $s^l = 3$ and $s^p = 0$ is 0.6. Although the case is the worst, one may be confused by the 0.6 deviation score and misunderstand that the prediction has a certain degree of correctness.

Reviewer dJ322024-08-09

I appreciate the thorough responses to my questions, the revisions made to the paper, and the additional experiments detailed in the "Common Concerns" section. As a result, I have adjusted my evaluation positively. From my current understanding, to better capture the properties of an image, multiple vectors are required instead of one. Within this set, some vectors carry a significant amount of the image's information, while others carry less. When assessing the similarity between two images, as their differences increase, the dot product of all the corresponding vector pairs will decrease, but those of vectors that carry more substantial information about the images will decrease faster. Thus, the relative scale relationship could change. Is my understanding accurate?

Authorsrebuttal2024-08-09

We sincerely thank you for your positive evaluation. Your interpretation on the "6-vector rationale" is insightful and has given us great inspiration: it triggers a deeper thinking on the underlying mechanism and raises a very plausible explanation. We will also continue our study, and if we find anything new, we will let you know. Thanks again.

Reviewer 2Ky27/10 · confidence 4/52024-07-15

Summary

This paper proposed a new image copy detection model for diffusion models. A dataset of replication labels (0 to 5) is collected using human labelers and is used to train a model for replication grading. The proposed model estimates a pdf of replication labels and minimizes the error with a pdf representation of class labels. The problem is timely, and the proposed solution is novel.

Strengths

A new model and framework for image copy detection (ICD) is proposed in the paper. The paper is well-written, and the results indicate a significant step forward compared to the mentioned literature.

Weaknesses

* The use of continuous labels 0-1 will be more useful in practice. * Some details of the labeling procedure are missing: how many labels per image, variance for each labeler, and across labelers. * The same dataset D-Rep is used for evaluation.

Questions

* I don't see why the authors did not choose continuous labels instead of the 6-categories approach. Instead of using argmax in eq(10), a weighted average score that is normalized between 0-1 will be more appropriate for practical applications. * Some details of the labeling processing are missing: how many images did each labeler label? What is the variance between the labelers for a specific image? * Evaluation can be more robust using a separate dataset with a different prompt/image generation procedures.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

Yes.

Reviewer 2Ky22024-08-10

Thanks for your reply and glad to see that my suggestion improved the original performance. It would be helpful to include the details you mentioned in Q2 in the paper.

Authorsrebuttal2024-08-10

We sincerely thank you again for your suggestion and will include the labeling procedure in A2 into the camera-ready version.

Reviewer vdEZ2024-08-11

Response to Authors

Thanks the authors for provding such detailed rebuttal. Most of my concerns are well-addressed. However, I highly suggest the authors to largely polish the writing and structure of the current manuscript in the revision. Page limitation exists for most of the publications. Yet, a paper should be complete by itself, and allows the readers to understand necessary preliminaries, methods, and experiments by only reading the main paper. I found the current manuscript has not reached this standard yet. Nevertheless, this is a good work with valid motivation. I would like to accept this paper for the task it studies. Therefore, I raised my ratings after rebuttal.

Authorsrebuttal2024-08-11

We sincerely thank you for your positive rating and will definitely move the required information to the camera-ready version (which has one more page) and polish our manuscripts accordingly.

Reviewer 2jDa2024-08-11

Response to Authors

I have checked the feedback from my fellow reviewers as well as the corresponding rebuttals. The concerns appear to have been resolved satisfactorily. My own concerns have also been addressed effectively. Therefore, I raise my score to 7 and confidence to 5. Thanks for authors’ efforts.

Authorsrebuttal2024-08-11

We are glad that our efforts have satisfactorily addressed the concerns raised and thank you very much for your thorough review.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC