Thanks for the insightful comments. Here, we carefully clarify each issue mentioned by the respected reviewer.

Q1. What is the motivation for conceptualizing real to fake as a progressive transition? Why it should be continuous rather than discrete?

R1. Thanks for this concern. We endeavor to address this with a systematic and comprehensive response:

• Motivation of Progressive Transition: Our paper seeks to effectively utilize both DF and BF data by taking the SBI and CBI as the intermediate anchors, where we conceptualize "Real->SBI->CBI->DF" as a process of getting more and more fake. It is intuitively apparent (Fig. 2 and Fig. 5 in the manuscript) and could be emphasized as accumulating forgery attributes (i.e., blending clue, ID-inconsistency, and generative artifacts). Hence, we leverage this 'inductive bias' to organize the latent-space distribution, and thus achieve utilizing both DF and BF data effectively.
• Continuous or Discrete: Building continuous bridges between adjacent anchors is based on the inherent natures and realizations of SBI, CBI, and DF. Here, we would like to illustrate these natures in detail.
  • Taking Real->SBI as an example: SBI is generated by blending two versions of one Real transformed by augmentations (e.g., blur, compression, and landmark shifting) with different severity levels. It can be emphasized that the unalignment levels of two transformed versions are controlled by an abstract parameter $\epsilon$, which represents the severity level of the blending clue. Blending with a small $\epsilon$ produces an image that can be treated as Real with little perturbations, while large $\epsilon$ produces an SBI with severe blending clue. Notably, we cannot assert a certain value of $\epsilon$ that when below it, the transformed image is fake and above it is real. That is, we cannot define a clear decision boundary between Real and SBI. Therefore, we can conclude that the transition between Real and SBI is Continuous according to the increase of blending clues.
  • Analogously, SBI to CBI might controlled by the increase of ID-inconsistency, and from CBI to Deepfake might controlled by the generative artifacts. Both can be subsequently treated as continuous transitions.

Therefore, we conceptualize the process of 'Real->SBI->CBI->DF' as a progressive and continuous transition. The necessity of continuity and progress, as we analyzed above, explains the rationale behind deploying transition and bridging. That is, the motivation to apply bridging and transition loss to simulate the conceptualized continuous and progressive transition.

Q2&3. Lacks experimental validation to the claimed motivation, that is, the latent space of VHT is not organized.

R2&3. Please refer to Common Responses R2, where we carefully address this concern.

Q4. The successful results in Face X-ray and SBI have already addressed the issue that we raised about substituting DF with BF.

R4. Thanks for your thoughtful consideration. We would like to clarify it as follows.

• As mentioned by the reviewer, the results of Face X-ray and SBI empirically show that BF may substitute DF considering 'DF-only < VHT< BF-only'. However, it is precisely the 'stereotype' that our paper actually seeks to break down.
• The contribution and conclusion of our work is: using both of them to organize the latent space progressively can be more effective than using BF only, and thus BF cannot substitute DF. We show detailed analysis and validation for this point in R2 of the Common Responses.
• This new answer is validated by the results of 'DF-only < VHT < BF-only < Ours' (Tab. 2 in manuscript).

Q5. $A^T$ is not defined

R5. Thanks for the kind mention. The superscript $T$ is commonly used to represent transpose matrix. We will explicitly define $T$ in the updated manuscript.

Q6. The rationale behind augmented with noise should be mapped to a more fake distribution

R6. We are grateful for your insightful concern. The transition loss is applied to simulate the progressive transition between two adjacent anchors. Specifically, we leverage 'noise+swallow network mapping' to build a less complex mapping path between adjacent anchors, simulating the path of progressive transition. Therefore, augmenting noise to feature actually represents the transition relationship between two adjacent anchors. Whereas in the case of Real->BF->DF, it represents getting more fake.

Q7. More datasets (WDF and DF-1.0) should be included.

R7. We have provided more experimental results including the datasets you've mentioned. Please refer to Common Responses R1.

Q8. Robustness comparisons with SoTA and multi-intensity.

R8. Thanks. In Author-Rebuttal-PDF-Fig. 1, we hereby provide robustness evaluation with multi-level unseen perturbations and comparing with SoTAs. The robust performance of our method indicates its effectiveness in complex real-world scenarios. We will update our manuscript with this more comprehensive robustness evaluation.

Q9. In Table 4, is R2B2D the same as D2B2R?

R9. Thanks for this insightful concern. From the perspective of implementation, the code of R2B2D should be the same as D2B2R except for inverting the attribute labels. Therefore, their experimental results are also expected to be consistent. Theoretically, D2B2R should be considered as the opposite of R2B2D, that is, the conceptualized progressive transition is getting more and more REAL.

Q10. What is the real/fake label of the mix of F_r and F_s?

R10. Thanks for your concern. As illustrated in Eq. 4 and line 172 of our manuscript, the label after bridging is assigned based on the mixing ratio $\alpha$. For example, the labels for $F_r$ is 0 and for $F_s$ is 1, and they are mixed with $\alpha=0.3$, the label for their mixed results should be $0.3×0+0.7×1=0.7$.

Dear Reviewer 6DHe,

Thank you for your thorough evaluation of our work. We are committed to incorporating your feedback comprehensively into our revised paper to improve both its content and overall quality.

As the discussion period draws to a close, we hope our response has sufficiently addressed your concerns. If there are any additional issues or points that require further clarification, we are more than willing to address them promptly.

Best regards, The Authors

Response to Reviewer Ssp1

We are thankful for the reviewer's positive comments and interest in our research. We hope that the following point-by-point responses will enable the respected reviewer to further recognize our work.

Q1. The argument of 'blendfake data is sufficient' is based on empirical observations on certain datasets, which may be not the universe.

R1. We appreciate the reviewer's insightful observation. Our main point is that while BF-only might not always outperform DF-only across all datasets, our method still demonstrates non-trivial advantages.

• To comprehensively investigate the effectiveness of DF and BF (using more datasets, not limited to "DF-series"), we enlarge our original evaluations by using 9 distinct deepfake datasets (see Common Responses R1). The results show that the situation of 'BF-only<DF-only' indeed sometimes exists, which validates the scoping concern from the respected reviewer.
• However, this does not undermine the significance of our method. Namely, it can be observed that the VHT still fails to consistently achieve the effect of "1+1>2" and shows degradation compared to BF-only. In contrast, our method effectively leverages both DF and BF data and achieves superior performance compared with VHT, and so does DF-only and BF-only.

Q2. More advanced and in-the-wild datasets (FakeAVCele, DFD, and RWDF-23) are recommended to be included.

R2. We have provided more experimental results including the datasets you've mentioned. Please refer to Common Responses R1.

Q3. The progressive transition pipeline could be promoted to be more generalizable, for instance, by incorporating compression.

R3. We appreciate the reviewer's interest in gaining deeper insights into the proposed progressive transition pipeline. We value this suggestion very much and carefully conduct an experiment based on the pipeline of 'Real->Real (Compression)->SBI->CBI->DF'. Notably, we do not place the Real (Compression) after DF as mentioned by the respected reviewer since we believe Compression is not a faker version of DF, considering it has no id-inconsistency or DF artifacts. Instead, Compression can be closer to the SBI, which actually involves compression operations in its creation process. The frame-level AUC results are shown below:

It can be observed that adding compression exerts little effect on experimental results, which may be due to the fact that random compression is already deployed as a data augmentation strategy for all training data. Still, we believe that this insight from the respected reviewer is of great value, which is very inspiring to us. For example, we may incorporate "Facial Beautification Algorithm" or "Photoshop Image" into the progressive pipeline to enhance the generalization and application scope. We are passionate about validating these ideas in our future works.

Dear Reviewer Ssp1,

We are encouraged by your thoughtful feedback and grateful for your interest in our research.

As the discussion period coming to a close, we would like to inquire if our response has adequately addressed your concerns. If there are any additional issues or points that require further clarification, we are more than willing to address them promptly.

Best regards, The Authors

Response to Reviewer t2ao

Thanks for your comments. Below we provide a point-by-point response to address the concern from the respected reviewer.

Q1. The attribution of the unorganized latent-space distribution lacks comprehensive experiments.

R1. Thanks for the concern. Please refer to Common Responses R2, where we comprehensively clarified this concern.

Q2. Writing issues

R2. Thanks. Please refer to Common Responses R3.

Q3. Results in Tab. 3 may be opposite to our statement in line 230.

R3. Thanks for the concern. Actually, in line 230 we stated: 'VHT is outperformed by BF-only', which is consistent with the experimental results in Tab. 3, that is the reviewer mentioned 'VHT is lower than that of BF-only'.

Q4. Concerns about the training dataset for different methods

R4. Thanks for your thoughtful consideration. Here, we redesign the structure of the table to clarify our training data settings. All results are extracted and summarized from the corresponding parts in Tab. 1, 2, and 3 in our manuscript. For fair comparisons, all methods are deployed with the same backbone, i.e., EfficientB4.

Our paper holds the view that naively hybridizing more training data may undermine the cross-dataset performance (validated by BF-only > VHT). Therefore, the difference in training set is not a biased setting, instead, it demonstrates the superiority of our method in leveraging multiple data effectively.

Dear Reviewer t2ao,

We deeply appreciate your dedicated efforts and insightful concern regarding our manuscript.

With the discussion period coming to a close, we would like to inquire if our response has adequately addressed your concerns. If there are any additional issues or points that require further clarification, we are more than willing to address them promptly.

Best regards, The Authors

Response to Reviewer LWpa

We sincerely appreciate the reviewer's positive comments and rating on our paper, and the following are our point-to-point responses

Q1. Choice of Blend Algorithms： Why apply SBI and CBI instead of other Blendfake methods?

R1. Thanks for your thoughtful suggestion. Actually, the terms SBI and CBI in our paper do not refer to two specific methods (i.e., SBI'22 [1] and X-ray), but a full set of methods with the techniques of Self-Blended and Cross-Blended. Therefore, we can surely take other Self-Blended methods or Cross-Blended methods in our current framework. To validate experimentally, we take FWA (another Self-Blended method) and I2G (another Cross-Blended method) to conduct a concise ablation study:

It is evident that our method can also enhance the performance of FWA&I2G trained VHT model, while using SBI'22&X-ary performs essentially better.

Q2. The reason for applying interpolation on feature-level, and could mixing parameters $\alpha$ are set multiply for enhanced performance.

R2. We genuinely appreciate the consideration.

• For the first question, the latent-space interpolation is performed to achieve feature bridging between two adjacent anchors (e.g., from SBI to CBI). This simulates the progressive and oriented transition of data becoming "more and more" fake starting from real. This process is illustrated intuitively in Figure 2 of the manuscript.
• For the second question, $\alpha$ is randomly sampled from $U(0,1)$, where $U(a,b)$ denotes the Uniform distribution with the bounds of $a$ and $b$. Therefore, for diversity and enhanced performance, $\alpha$ is indeed a parameter with multiple values in our method.

Q3. Possible Typos

R3. Thanks. Please refer to Common Responses R3.

[1] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR2022

Dear Reviewer LWpa,

We greatly appreciate your careful review and the insightful comments you shared regarding our manuscript.

With the discussion period coming to a close, we would like to inquire if our response has adequately addressed your concerns. If there are any additional issues or points that require further clarification, we are more than willing to address them promptly.

Best regards, The Authors