Review Reports

Round 1

Reviewer 1 Report

The article is written in a very good quality, from the beginning it gradually introduces the reader to the issue of Thangka artworks, the importance of its preservation and the ways of creating these works in an artificial way. By stating the challenges associated with the creation of Thangka illustrations accurately from the early chapters, the intent for addressing these challenges is also precisely defined. The methodology is described excellently, and the various definitions are well explained and supported by appropriate equations.

Furthermore, common metrics are chosen to evaluate the necessary quality parameters of the generated images. The description of the training or the loss function itself is given due attention, where the weights of the different parts of the loss function are chosen without further description. I highly recommend adding a justification for the choice of the parameters/coefficients just mentioned. A study of the effect of these parameters on training and results would also be beneficial.

The results are well described separately in an appropriate chapter, however I would recommend an extension of the conclusions where a more comprehensive summary of the results obtained would be given. Future work could also be discussed, as despite the impressive results there is room for improvement of the technique.

I identified only a small number of typos in the text, namely on line 137, where a space after the encoder is missing, and in Figure 4, in which the word "ferture" is written on instead of "feature" in all occurrences. In addition to these errors, there is an error in Figure 3 where the arrow (AdaIN module connection) is missing in the first CAR module if they are supposed to be identical. In Figure 4, I recommend to unify the occurrence of arrows "into the void". 

The references include articles in the form of preprints. I'm not sure if such can be used for this type of publication and recommend rewriting them into published versions.

Author Response

Response to Reviewer 1 Comments

We sincerely thank you for raising these good questions.

 

Point 1: The description of the training or the loss function itself is given due attention, where the weights of the different parts of the loss function are chosen without further description. I highly recommend adding a justification for the choice of the parameters/coefficients just mentioned. 

 

Response 1:

• We reorganized Section III (3.5 Reconstruction Loss) on page 6 (lines 195-200) to further describe the reason for our choice of this loss function. The revision helps the reader to have a clearer and deeper understanding of our choice of approach, rationale and advantages.
• We reorganized Section III (3.5 Adversarial Gen Loss) on page 6 (lines 202-216) and added some explanations to Eq. (2) to more clearly express the intent of Eq. (2) and make our paper easier to read.

 

Point 2: The results are well described separately in an appropriate chapter, however I would recommend an extension of the conclusions where a more comprehensive summary of the results obtained would be given. Future work could also be discussed, as despite the impressive results there is room for improvement of the technique.

 

Response 2: We expand our conclusions on page 12 (lines 331-339) to include possible future related work that summarizes our conclusions more comprehensively. The reader can now develop a more comprehensive and in-depth understanding of our paper.

 

Point 3: I identified only a small number of typos in the text, namely on line 137, where a space after the encoder is missing, and in Figure 4, in which the word "ferture" is written on instead of "feature" in all occurrences. In addition to these errors, there is an error in Figure 3 where the arrow (AdaIN module connection) is missing in the first CAR module if they are supposed to be identical. 

 

Response 3: We corrected all typos and tiny mistakes the reviewers mentioned.

Point 4: In Figure 4, I recommend to unify the occurrence of arrows "into the void".

Response 4: We unified the arrows in Figure 4 on page 6.

Point 4: The references include articles in the form of preprints. I'm not sure if such can be used for this type of publication and recommend rewriting them into published versions.

 

Response 4: We revised most of the arXiv references, improved online references, and added some newest literatures in this version.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper presents a smart colorizing method based on deep learning techniques. The paper focuses on two datasets: Danbooru and Thangka and the methodology is quite good, but the paper should be improved.

Danbooru dataset is public while Thangka is not, I think it would be useful to have it also public or at least available for reviewers. On the other hand, the networks are high level described but there is no code for results replication. Usefulness of this paper can be improved if Thangka dataset would be made public along with the code.

Author Response

Response to Reviewer 2 Comments

We sincerely thank you for raising these good questions.

 

Point 1: Danbooru dataset is public while Thangka is not, I think it would be useful to have it also public or at least available for reviewers.

 

Response 1: We mentioned in the Data Availability Statement on page 12 (lines 349-350). Currently, our related work is still ongoing in our laboratory, and the Thangka datasets used in this study will be publicly available later. If needed, they can now be provided by livingsailor@gmail.com.

Point 2: On the other hand, the networks are high level described but there is no code for results replication. Usefulness of this paper can be improved if Thangka dataset would be made public along with the code.

Response 2:

We sincerely thank you for your attention to the Thangka dataset and the codes. The relevant data sets and code will be released later. We welcome other researchers to discuss with us.

Author Response File: Author Response.pdf