CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting

Thank you to the reviewer for pointing out inaccuracies and offering suggestions. We’re glad the work is considered interesting and easy to follow. We hope our responses and revisions will improve clarity for future readers.

W1 [More comparisons]:

We thank the reviewer for the suggestion. We conducted additional comparison with SGSST (CVPR 25) and ABC-GS (ICME 25) (for 3D) as well as ViSt3D and ReReVST (for video). Moreover, we have included results obtained using CLIPGaussian with an alternative set of hyperparameters (lambda_p​=35, feature_lr = 0.002), which produces slightly lighter stylization. We call this variant CLIPGaussian-Lite. As a 4D style transfer is a new field, we already compared CLIPGaussian with all relevant baselines. The updated results for 3D and Video are shown in the tabs below.

TABLE xZpT.A

CLIPGaussian, in both the base and light variants, outperforms SGSST and ABC-GS in stylization quality metrics (CLIP-S and CLIP-SIM). Notably, CLIPGaussian-Light achieves the highest scores across all consistency metrics. It is important to highlight that G-Style significantly increases the model size; typically by more than a factor of two, whereas CLIPGaussian offers competitive results without altering the underlying model size.

TAB. xZpT.B

CLIPGaussian also outperforms ViSt3D and ReReVST. Additionally version with lighter stylization has the best consistency while still outperforming baselines.

W2 [Lack of sufficient quantitative evaluations]: Please find quantitative evaluation for 3D in: Tabs. 1-12, 4D: Tab. 13, Video: Tabs. 14, 15. However, as suggested by reviewers 8XY8 and XWe5, we conducted additional quantitative evaluation using flow-based consistency metrics for 3D, Video and 4D as well as CLIP-based consistency metrics for 4D.

TAB. xZpT.C

TAB. xZpT.D

In response to the reviewer XWe5, we evaluate consistency of CLIPGaussian in 4D setting (see Tabs. XWe5.B-D).

It is important to note that the interpretation of style transfer metrics is not trivial. CLIP-S and CLIP-SIM scores would be maximized if the generated object were identical to the style image or prompt, which would clearly be an undesirable outcome. On the other hand, consistency metrics are maximized when the output is identical to the original object, which also indicates poor stylization. This trade-off can be easily seen in Tab. 11. We believe that CLIPGaussian strikes a meaningful balance between these two objectives, achieving high-quality style transfer without compromising the consistency of the object or video. Another important factor, which is not measured by any metrics in universality. All other methods work on only one medium (3D, 4D, 2D or video) and support only one modality (text or image). CLIPGaussian is the first method that works in all of these settings while being either superior or at least comparable in quality.

In case of user study, we evaluated “quality” or “visual appeal” in the Question 2: "On a scale from "Very Low" to "Very High", how would you rate the visual appeal of each generated result?”. We will highlight it in the revised version of the main manuscript.

Q0 Main contribution of the paper

Our main contribution is a unified framework for various data modalities such as 2D, 3D, 4D, and videos etc. This broad applicability was positively highlighted by the reviewers (e.g., Rev. ed8o and Rev. XWe5). Our framework demonstrates competitive performance across modalities, which can be especially seen in the case of 3D and 4D, where it outperforms existing approaches. We acknowledge that our performance is not yet on par with current state of the art specialized approaches in the case of images, which are often based on large diffusion models or ChatGPT-like architectures. In contrast, our method is the first to propose a unified multimodal GS based style transfer approach.

Q1/ Q2 Motivation of each component and its effect

CLIPGaussian operates by directly optimizing gaussians (either in 3D, 4D, Video or Images) using losses based on CLIP embeddings. Analysis of loss terms can be found in the Appendix B.2.3 (lines 722-736). Description of the losses can be found in Section 3 (lines 217-252). The loss consists of four main parts:

1. Content Loss (Eq. 2): This is a classical loss computed using VGG-19 embeddings. During each optimization step, it encourages the generated view to preserve the content of the ground truth. In the case of video, it ensures that each generated frame retains the original semantic content; e.g., that the subject remains a camel or a bear (see Figs.12 and 13).
2. Directional CLIP Loss (Eq. 3): This component is responsible for global style transfer. It compares the CLIP embedding of a rendered view with that of the corresponding original view, both conditioned on the style prompt or style image. This encourages the overall appearance of the object to align with the desired style.
3. Patch CLIP Loss (Eq. 4): This loss is equivalent to the Directional loss but is applied to random patches of the rendered view rather than the entire image. It ensures that each patch of the rendered view is stylized. This loss is responsible for the texture of the stylized object, as it operates on smaller fragments.
4. Background loss (bg loss): This loss is applied only for 3D and 4D objects. In these cases, we aim to stylize the object itself without affecting its bg. For example, this loss ensures that the white bg behind the Lego set in Fig. 6 remains unchanged. In contrast, StyleGaussian does not enforce bg consistency on the same image.

W3 [Ablation studies regarding the loss terms]: We have already conducted an ablation study in the context of 3D stylization. The results are presented in Fig. 24 and Tabs. 12 and 13 of the manuscript. The parameters lambda_d and lambda_p correspond to the weights of their respective loss terms (directional loss and patch loss). In both cases, we examined the effect of setting these weights to zero. The effect of the bg loss can be seen on Fig. 26. We will highlight this fact in the manuscript. We believe that this will improve understanding of losses.

Q3 [The differences between previous methods will be needed] Instruct-GS2GS:

CLIPGaussian is not a unified version of Instruct-GS2GS (I-GS2GS). While I-GS2GS stylizes training images using a diffusion model before fine-tuning the Gaussian splats, CLIPGaussian directly optimizes the Gaussian representation using CLIP-based losses. This allows high-quality style transfer with both image and text conditioning, unlike I-GS2GS, which supports only text prompts. Additionally, CLIPGaussian does not rely on any pretrained diffusion models and is applicable across modalities, whereas I-GS2GS is limited to 3D. A direct comparison with I-GS2GS is provided in Fig. 19 and 20. We will add a description of I-GS2GS to the related work section. Thank you for pointing out this omission.

L1 [Social Impact]:

ClipGaussian enables cross-modal generality as a plugin based on GS models. Using CLIP as a multimodal vision and language model motivated it to enable high-quality stylization and expanding creative possibilities. Gaussian based representation enables the creation of high-quality real-time renders which is a key challenge in computer vision, especially in broad applications in AR, gaming, and digital content creation. We will add this to manuscripts.

We are grateful for your careful review, which contributed to improving the clarity of our paper. If anything remains unclear, we would be glad to answer any questions.

Thank you for your very insightful review.

Q4/W2 [hallucinated human-face-like structures]:

Thank you for this very interesting question. The appearance of human-face-like structures in the output is an expected outcome, resulting from the combination of the chosen patch_size and the specific style prompt.

For example:

• in Fig. 10 (bottom row), the scream-face-like patterns are a natural result of the characteristic visual elements in the refereed image "The Scream by Edvard Munch".
• in Fig. 11 (top row) the prompt “Beksinski painting” caused the creation of human-like structures. This is consistent with Beksinski’s style, known for dark landscapes and ghostly, human-like characters.

In both cases, these structures are not hallucinations but rather stylizations aligned with the chosen references. Thepatch_size parameter controls the local stylization. For example for patch_size=128, each patch of size 128x128 is stylized. For this reason, detailed repetition can be controlled through the patch_size param. This effect is illustrated in Fig. 23 (Lego scene). Increasing thepatch_size results in more star-like structures caused by the influence of the prompt “Starry Night by Vincent van Gogh.”

We will add a clarifying note to the experimental description

Q1/W3:

• 2D: The paper compares only two methods...

We compare the performance of our model on 2D images with AdaIN and StyTr² in Fig. 10 and 31, with InstructPix2Pix and ChatGPT (DALL·E) in Fig. 11, and with CLIPStyler and FastCLIPStyler in Fig. 30. In total, we compare our model against six other methods, not just two. Moreover, in the manuscript, we explicitly acknowledge that, given the highly developed state of 2D image stylization methods, our results in 2D may not match the visual quality achieved by large or diffusion-based models (lines 358–359). AdaAttN (2021) is an older method and can no longer be considered state of the art. While we do not directly compare to "Style Injection in Diffusion", we do evaluate our model against ChatGPT (DALL·E), which is also based on diffusion.

• 3D: No comparisons are provided against relevant works such as SGSST (which is directly related), StylizedNeRF, Style-NeRF2NeRF, or MM-NeRF.

Our goal was to compare style transfer on fully trained 3D objects. NeRF-based models have significantly higher computational requirements and cannot be used in the same setting as Gaussian splats. For these reasons, we decided to compare CLIPGaussian only with other GS-based techniques. We emphasize that the base models were identical across all evaluated methods (with the exception of Instruct-GS2GS, which was incompatible). However, in response to the reviewer's suggestion, we additionally compare our model against two recent baselines: SGSST (CVPR 2025) and ABC-GS (ICME 2025). Moreover, we have included results obtained using CLIPGaussian with an alternative set of hyperparameters (lambda_p​=35, feature_lr= 0.002), which produces lighter stylization. We call this variant CLIPGaussian-Lite. The updated results are shown in the table below.

TABLE 8XY8.A

CLIPGaussian, in both the base and light variants, outperforms SGSST and ABC-GS in stylization quality metrics (CLIP-S and CLIP-SIM). Notably, CLIPGaussian-Light achieves the highest scores across all consistency metrics. It is important to highlight that G-Style significantly increases the model size; typically by more than a factor of two, whereas CLIPGaussian offers competitive results without altering the underlying model size.

• Video: ...no comparison to methods such as AdaAttN, StyleMaster, ViSt3D, MCCNet, or ReReVST.

Since MCCNet (2020), ReReVST (2020), and AdaAttN (2021) are relatively older methods, we chose to compare our model with more recent baselines. We acknowledge that StyleMaster would be a valuable baseline; however, as its code was unavailable at the time of the submission, it was not possible to include it in our comparison. Nevertheless, we will add them to the related work section. In response to the reviewer's suggestion, we additionally compare our model with ReReVST and ViSt3D using the same dataset and experimental setting as described in the manuscript. The updated results are shown in the table below.

TABLE 8XY8.B

• 4D:...image-guided stylization results are missing, and comparisons with StyleDyRF, Instruct 4D-to-4D, etc., are not discussed

Instruct 4D-to-4D has been discussed in main manuscript lines 114-121 in section Related Works and line 315-318 in section Experiments. Instruct 4D-to-4D focuses mainly on the color palette, while our approach additionally modifies the geometry of the scene. Consequently, the results obtained by CLIPGaussian contain more details, which in our opinion better resembles the referenced style. We will add this discussion. The supplementary also includes image-guided results for 4D, see Fig. 25. We will add more examples to the main manuscript.

We thank the reviewer for pointing out the StyleDyRF method, which is a relevant NeRF based zero shot style transfer approach that stylizes dynamic scenes represented by Dynamic NeRF. In contrast to our method, it supports only image based style conditioning. In our work we compared against more recent img-conditioned zero shot approaches such as 4DStyleGaussian, which similarly to our method works on a GS scene representation. We will add the discussion to the related work section.

Q2/W5 [optical flow-based metrics for video]: We conducted an additional evaluation of CLIPGaussian using short-range and long-range consistency metrics, following the implementation and settings from StyleRF. Specifically, we warp one view to another using optical flow and softmax splatting, and then compute the masked RMSE and LPIPS scores to assess stylization consistency. For short-range consistency, we average the scores over pairs of the i-th and (i+1)th frames; for long-range consistency, we use pairs of the i-th and (i+7)th frames. We evaluate Video-CLIPGaussian and the baselines on the same dataset used in the paper. The results are presented in the table below.

TABLE 8XY8.C

The interpretation of these metrics is not trivial. Consistency metrics are maximized when the output is identical to the original object, which suggests poor stylization. Notably, only Text2Video deviates significantly from the other methods in terms of consistency. It is also worth noting that while Rerender achieves low metrics values, the stylization and content of the videos decreases in time (see suplementary_material/Video/comparisons/comparison_white_wool.mp4). CLIPGaussian offers superior stylization quality compared to the baselines, and we believe it achieves high-quality style transfer without compromising the consistency of the object or video. Additionally in the response to the reviewer XWe5, we evaluate 3D spatial consistency (Table XWe5.A) as well as 4D spatial and temporal consistencies using the same metrics (Tables XWe5.B-D).

Q3/W4 [user study]:

The user study is described in detail in App. B.2.2. It was conducted carefully using the CLICKworker platform, which enabled us to recruit a balanced group of participants (lines 641-644). The structure of the survey closely follows that used in the G-Style paper[21], including the preference ranking question described in line 660. The questions were constructed identically to the reference work. We additionally proposed question 4 (line 674) and additional question (line 680). This alignment allowed us to make meaningful comparisons while building on validated recent prior work. We would like to emphasize that the user study is a supporting evaluation, not the core of our contribution. The primary contribution of our work lies in presenting a general and flexible stylization framework, which can be applied to any Gaussian-based model and supports both text- and img-driven conditioning. To our knowledge, no other existing method offers this level of adaptability

Q5, Q6/W1, MW1[Paper formatting]: Thank you for highlighting the editorial corrections and typos we will address them carefully.

We would be happy to address any further questions and provide clarifications.

I thank the authors for the detailed and constructive rebuttal.

Content Fidelity and Stylization Hallucinations (Q4/W2):

The clarification on the appearance of face-like patterns (e.g., due to patch size and the semantic content of the style prompt) was helpful. I now agree these are not hallucinations in the negative sense but a result of stylization driven by strong structural priors in the reference styles. The authors' suggestion to add explanatory notes in the manuscript would improve clarity.

---

Cross-Modality Evaluation (Q1/W3) and Temporal Consistency (Q2/W5): While the rebuttal adds comparisons across 2D, 3D, video, and 4D modalities, I remain concerned about the evaluation in video stylization, especially for real-world, high-motion scenarios.

Specifically:

• Most video stylization experiments are conducted on low-motion datasets (e.g., DAVIS) or synthetic datasets like D-NeRF, which feature clean geometry, low texture variance, and minimal deformation.

• These settings do not reflect real-world stylization challenges such as background flicker, motion blur, and viewpoint inconsistencies.

Regarding temporal consistency evaluation, I also note that the entire video frame—including both foreground and background—should be considered during optical-flow-based evaluation. Restricting evaluation to foreground regions only (as implied by masked metrics in 3D/4D settings) underestimates flickering and inconsistency, which commonly arise in real-world backgrounds.

The temporal consistency metrics (short- and long-range LPIPS and RMSE) reported by the authors are informative but insufficient to assess long-term perceptual quality in high-motion scenarios:

• LPIPS, while good for short-term perceptual similarity, cannot capture motion-induced artifacts across large frame gaps.

• RMSE is sensitive to natural motion and may penalize valid stylization changes, thus leading to misleading conclusions.

I appreciate the inclusion of baselines such as ViSt3D and ReReVST. However, methods like AdaAttN, although older, show evaluation in motion of the video contexts and could offer a valuable comparison reference for robustness assessment. Moreover, while CLIPGaussian-Lite shows improved consistency scores, its stylization quality remains unverified as no visual results are provided in the paper or supplementary. Given the stylization-consistency trade-off, this is a non-trivial omission.

---

Q3 / W4: User Study

While the authors clarify their user study setup and note the influence from G-Style, it is still limited in scope (3D domain only) and uses a less interpretable rating-based approach. The lack of forced-choice comparisons and the absence of study results across other modalities reduces the significance of this evaluation.

---

Overall

The paper introduces a promising universal stylization framework with clear strengths in modularity and multi-modality support. However, I believe that the video modality, which is central to the paper’s "universal" claim, is still under-evaluated, particularly in challenging, real-world, high-motion scenarios.

I would be willing to raise my score if the following are clearly addressed in the rebuttal and supplementary material:

• Reporting optical flow-based metrics for temporal consistency and direct comparison with baseline video stylization methods (e.g., AdaAttN, ViSt3D, ReReVST).

• Inclusion of visual results and consistency metrics on real-world, high-motion videos, using datasets beyond DAVIS and D-NeRF. If authors commit to doing it in a camera-ready revision.

• Presentation of CLIPGaussian-Lite visual outputs. If authors commit to doing it in a camera-ready revision.

Unless the above is satisfied, I will maintain my current rating.

---

Thank you to the authors for their detailed follow-up and appreciate the additional clarifications provided.

Q1: Temporal Consistency and Optical Flow Evaluation

I thank the authors for incorporating optical flow-based metrics, following the ViSt3D setup, and for reporting performance against baselines such as AdaAttN, ViSt3D, and ReReVST. Based on the numbers in Tables 8XY8.B2 and 8XY8.C2, the temporal stability of CLIPGaussian-Lite is now much clearer and quantitatively substantiated. The results on the DAVIS dataset show that CLIPGaussian-Lite consistently achieves favorable frame-to-frame consistency and competitive performance across both short and long temporal windows.

I am assuming that the authors have used stylized video results, and there might be a typo in the mathematical description of the FoF metric, where "Style" should likely be corrected to "Stylized" in the optical flow equation.

Q2: Evaluation on Real-World, High-Motion Videos

I also appreciate the authors’ acknowledgement of the importance of testing on real-world, high-motion sequences and their commitment to reporting results on the Ultra Video Group (UVG) dataset in the camera-ready version. This dataset, with its high frame rate and diverse content, is a strong candidate for validating video stylization models under realistic conditions.

Additionally, the use of PanopticSports for 4D stylization evaluation is a valuable step forward. The inclusion of consistency metrics for different prompts (Table 8XY8.D) improves the empirical grounding for 4D results.

Remaining Concerns

While the authors have made credible commitments for extended evaluations and visual results in the camera-ready version, the current rebuttal still lacks:

• Rich qualitative visual analysis due to NeurIPS 2025 media-sharing constraints.
• A sufficiently broad and standardized user study using forced-choice protocols.

These aspects limit the interpretability of stylization quality and user preference, especially across modalities.

Overall Evaluation

Based on the newly reported quantitative results, and the authors’ commitment to address visual and qualitative aspects in the final submission, I find the rebuttal satisfactory and convincing. I have accordingly improved my score to reflect this update.

We would like to thank the Reviewer for the kind words and interesting indication of gaps, we believe that the newly designed experiments will indeed help to improve the work.

**W1/Q1 $Metric of 3D Consistency$ **

As suggested, we evaluated the CLIPGaussian using short-range and long-range consistency following the implementation and settings from StyleRF. Specifically we warp one view to the other according to the optical flow using softmax splatting, and then compute the masked RMSE and LPIPS scores to measure the stylization consistency. For short-range we average over pairs of i-th and (i+1)-th frames and for long-range we take i-th and (i+7)-th frames. We evaluate 3D-CLIPGaussian and baselines on the same data as in the paper. Additionally, as asked by the reviewers 8XY8 and xZpT we evaluate against two more baselines, SGSST (CVPR 2025) and ABC-GS (ICME 2025). Moreover, we have included results obtained using CLIPGaussian with an alternative set of hyperparameters (lambda_p​=35, feature_lr = 0.002), which produces lighter stylization. We call this variant “CLIPGaussian-Lite”. It is important to note that this model still outperforms other methods in terms of stylization quality (CLIP-S and CLIP-SIM), with the exception of G-Style, which achieves slightly higher scores at the cost of more than doubling the model size (see table Tab. 8XY8.A in response to reviewer 8XY8).

TABLE XWe5.A: Consistency metrics for 3D

Additionally we also measure short-range and long-range consistency for video style transfer. The table can be found in response to reviewer 8XY8 (TABLE 8XY8.C).

W2/Q2 $Quantitative Evaluation of Consistency in 4D Style Transfer$ / Q3 $Spatial-Temporal Consistency$ :

CLIPGaussian is built on top of specialized architectures tailored to each modality. For 4D data, our framework is implemented on top of D-MiSo. As input, we use a fully trained D-MiSo model, which already addresses temporal consistency through a dedicated deformation network that models changes in the Gaussians throughout time. During stylization, we keep the weights of this network fixed to ensure that the stylization process does not directly affect temporal consistency. We assume that D-MiSo (or any other 4D model adapted for use with CLIPGaussian) handles this aspect. Nevertheless, in response to the reviewer's suggestion, we evaluate two types of consistency in 4D data:

1. The first is consistency across different views at a fixed moment in time. This is similar to the 3D setup, as described in Equations 11 and 12. To evaluate this GT is required for each view at a given time
2. The second is temporal consistency, where we fix the camera and analyze how the object changes over time.

In the supplementary material, we included visual examples that reflect both camera changes and temporal change for example:
supplementary_material/4D/examples/jumpingjacks_cubism.mp4

Unfortunately, the D-NeRF dataset does not provide GT across time for each camera. Only one time point is available in the data for each camera. However, we agree with the reviewer that in both cases, considering the metrics is very important. Therefore, we decided to conduct two experiments, supplementing Tab. 13:

1. For a set of time steps {0, 0.1, 0.2, ..., 1}, we used the test cameras from NeRF-Synthetic (200 views, as in the jumpingjacks_cubism.mp4) to generate reference images using the base D-MiSo model. Then, we generated the corresponding stylized images and computed the mean CLIP-F and CLIP-CONS metrics for time steps

TABLE XWe5.B:

From this experiment, we can see that it is difficult to indisputably determine which styling is best based on the selected metrics. We emphasize this in line 778 that visual comparison (Fig. 28) reveals substantial qualitative differences in output fidelity. Fig. 8 shows that with a larger batch, the shape of the object is better preserved (this is clearly visible for patch=32; where the mutant's head was more accurately reproduced only when we used batch=4 ). We will add additional metrics to the manuscript and describe our observations clearly.

2. For each test camera from the D-NeRF dataset, we generated stylized images at {0, 0.05, 0.1, ..., 1} time points. Next, we calculated the average metrics to evaluate temporal consistency across time for each camera.

TABLE XWe5.C

An interesting observation is that in almost every case, when using selected hyperparameters, we see an increase in metrics compared to the base model. In our case, a lot depends on stage 1; a good idea for future work would be to check the performance of colors represented using RGB instead of spherical harmonics in order to avoid light inconsistencies between views. We will include this observation in the manuscript as well.

TABLE XWe5.D

It is interesting that despite differences in the (well-represented) prompt, the metrics usually behave very similarly for the selected object and the changes are minor. This emphasizes the importance of object reconstruction (stage 1). Nevertheless, it also shows that our model yields consistent results across different prompts.

We sincerely thank the reviewer once again, as the new experiments provide additional insights into our model. We will include both the quantitative and qualitative results in the paper. To further highlight the effect of time and viewpoint changes, we will also include visualizations for additional objects, showing reconstructions and stylized versions at selected moments and camera positions.

We sincerely thank the reviewer for their thoughtful comments and interesting questions. We are especially pleased that the overall pipeline was well understood and that its contributions were clearly recognized. We hope the responses below will address any remaining concerns and help us further improve the quality of the paper.

W2, W3
$Inherited dependence on base GS model, Adding domain specific loss terms$ :

In this work we used background loss as an example of a domain specific loss term. It was only used in case of 3D and 4D objects to ensure that the background won’t be affected by the stylization. In future work, it would be a good idea to provide additional support for the base model, especially in terms of domain-specific consistency. We believe that adding the appropriate loss (for example time related in case of 4D/Video) would contribute to the correctness of the base model and stage 2: the styling itself. We will highlight this as a future work direction.

Q1 $Control over geometry$ :

Our model is based on the Gaussian Splatting representation. We can consider editing the object as in the case of the D-MiSo/VeGaS/MiraGe model (all of these base architectures support object editing), because the stylized representation fully retains the properties of the base model. When it comes to styling, we can modify the position as well as the scaling and rotation of corresponding Gaussian Splats. Hence, It is possible to achieve the effect of “detaching” the Gaussians from the reconstructed object. An interesting example is the stylized object using the text prompt Fur, see Fig. 1, Fig. 3, which can result in the effect of fluffy fur.

Another interesting prompt is cubism, which can be seen in supplementary materials under /supplementary_material/4D/examples/jumpingjacks_cubism.mp4.
We can see that the character's face has become more angular, typical to the style, and is obtained by changing the geometry of the head (face and hair).

We will add visual examples to the manuscript.

In our work, we focused on hyperparameters related to the base model, but in future work, it is possible to consider a loss responsible for controlling, among other things, the position of the gaussians, which would control the degree of freedom the splats are allowed to move in space; this would allow for greater control over stylization using prompts such as Fire.

Q2 $Failure cases$ :

It is difficult to say definitively that “the method can better imitate styles given by the text prompts (in comparison to reference images)”. We believe it strongly depends on the prompt/image and object/scene. For example, in our opinion the “mosaic” effect on the bonsai scene works better when it is conditioned by the selected image (Fig. 21) than when it is conditioned by (general) text “Mosaic” (Fig. 20). Nevertheless, an important element is how a given text is represented by CLIP. In general it is interesting how the model represents a given textual/visual concept.

Based on our experience we have identified three common categories of failure cases mostly related to text prompts

1. Too general concepts, less known concepts:

Since we are using CLIP representation, the concept should be well represented by CLIP embedding. We considered 3 prompts:

TABLE ed80.A

CosineSimilarity(prompt1, prompt2) = 0.802
CosineSimilarity(prompt1, prompt3) = 0.531
CosineSimilarity(prompt2, prompt3) = 0.382

We observed that for general prompts such as “Impressionism”, the model tends to fill the background, and hollow spaces (e.g. between fingers). This improves the quantitative results, but in our opinion it decreases the visual perception of the stylization. We support this claim using the CLIP-S metric, see the difference between lambda_bg=0 and lambda_bg=500 in Tab. ed80.A and Fig. 26 in the manuscript.

The concepts of “Impressionism” and “Painting by Claude Monet” are well represented by the CLIP model. Additionally, they are closely related, which is visually supported by the models trained using those prompts. On the other hand, concepts such as “Woman with(...)” are heavily influenced by details like an umbrella, which spoils the visual effect in this case. This pattern is shown by the CLIP-SIM metric. Due to the rebuttal rules we are not allowed to include visual comparison in our response, we will add it to the manuscript when possible.

2. Length and detail of general concept prompts:
We conducted an experiment using jumpingjacks in which we considered three prompt lengths:

TABLE ed80.B

We noticed that if a shorter prompt is considered, the CLIP model finds a certain representation of a specific word/short prompt related to a general concept. Creating a nice and consistent style, similar to Fig. 2. If a longer prompt is used, we can expect an averaged representation of the concepts found in the prompt, which may cause some inaccuracies in the stylized image. We will add a visual comparison to the manuscripts.

In both cases (1, 2), the styling emphasizes artifacts, especially in the background. We can mitigate the artifacts in the background using bg_loss and patch_size. This is particularly evident in detailed areas, e.g. on the hands (see Fig. 26, 27). But, visual assessment is very difficult and subjective.

3. Concepts that are difficult to represent, such as hope, country names, random or beauty.

We will add a comparison and additional experiments to the manuscripts.

Q3 $Diversity of reference images or text prompts$ :

For the user study, we prepared prompts and corresponding images, which we presented in Fig. 19, Fig. 20 (text prompts), Fig. 21, Fig. 22 (image prompts). The number of object-stylization pairs was prepared so that the average time to complete the survey was 10 minutes. A detailed description of the user study can be found in B.2.2. For evaluations based on text prompts, we wanted to avoid introducing visual bias, and we did not attach any reference images. In the context of “Starry Night by Vincent van Gogh” and “Scream by Edvard Munch”. We asked an additional question at the end of the survey: “Were you familiar with the following images before?” Users could answer “Yes” or “No.”

However, in our work, we indicate very different text prompts. In particular, supplementary_material/4D/examples/trex_many_styles.mp4 and supplementary_material/4D/examples/lego_many_styles. mp4, you can find over 26 different prompts; we agree that there are fewer for image conditioning (about 8; see Figs. 6, 8, 10, 12)

We will add this information to the main paper to improve the clarity of the work.

We thank the reviewer once again for feedback. We strongly believe that the additional experiments have further strengthened the contribution of our work. If there are any remaining questions or points requiring clarification, we will be happy to address them.