BrushGaussian: Brushstroke-Based Stylization for 3D Gaussian Splatting

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Review Report

1. Overall Assessment

The manuscript presents a novel style transfer method for 3D Gaussian Splatting that prevents local stylization of geometry by simulating brushstrokes. The texture compensation method with geometric transformation includes mechanisms for simplifying the scene through Gaussian aggregation. The authors argue the limitations of existing 3D style transfer methods (e.g. NeRF, focal point cloud), which are mainly based on global color and texture, and demonstrate how their approach allows for local artistic features, such as brush strokes. The method is well described, accompanied by results, comparative experiments, and multi-site evaluation (e.g. CLIP similarity), which generally provides a convincing argument. However, some aspects of the methodology need clarification, and it is also worth strengthening the discussion about the limitations of the method and its reproducibility.

2. Main Comments. Novelty and Contribution

• The authors position their method at the interface of geometric and texture stylization for 3D scenes. Unlike previous approaches (e.g. ARF or StyleRF), this work performs stylization at the level of individual Gaussians through texturing with local orientation in mind.
• The proposal to use scaled Gaussians as brush strokes is an interesting idea that could indeed have wide applications in visible styles.
• However, it would be worth emphasizing in more detail the differences from Chao et al. [76], who also work with textured Gaussians. The current explanation focuses only on the accuracy of the orientation, which can be strengthened by a direct visual comparison.

3. Results

• The visual results demonstrate the advantages of the proposed method. The authors have performed a comparison with ARF and StyleRF. The results in Figs. 7–8 look convincing.
• Table 1 effectively shows the comparison in numerical terms.
• The authors honestly mention the limitations associated with symmetry and texture flipping, as well as the need to manually configure additional clusters. It would be appropriate to add an example with an unsuccessful stylization to demonstrate these problems.
• The article is written competently, and the language is clear and technically correct. All terms are presented correctly, justified, and complicated. The style is formal, academic. The figures are informative, with appropriate captions.

4. Recommendation

1. The clustering process for Gaussian pruning is described, but it is worth providing more technical details: how is the number of clusters chosen? Has the impact of this number on the style result been assessed? Discuss the impact of many clusters on the stylization result.
2. Clarify the differences from recent works, especially Chao et al. [76], perhaps through additional visualization.

5. Conclusion

This article is an interesting and technically sound work outside the scope of 3D style transfer. It extends the application of Gaussian Splatting to artistically oriented tasks and demonstrates a strong experimental component.

After minor refinements, I recommend it for publication. Solution: Minor Revision (minor edits before publication)

Author Response

Comments 1: The authors honestly mention the limitations associated with symmetry and texture flipping, as well as the need to manually configure additional clusters. It would be appropriate to add an example with an unsuccessful stylization to demonstrate these problems.

Response 1: Thank you for your valuable suggestion. We agree that including failure cases can more clearly illustrate the limitations of our method. Accordingly, we have added Figure 9, which presents unsuccessful stylization examples caused by texture flipping and symmetry-related issues. Additional explanation has been provided on page 14, lines 407–418, to further elaborate on the observed artifacts and their potential causes. As for the issue regarding manually configured clusters, we will discuss it in Response 2.

 

Comments 2: The clustering process for Gaussian pruning is described, but it is worth providing more technical details: how is the number of clusters chosen? Has the impact of this number on the style result been assessed? Discuss the impact of many clusters on the stylization result.

Response 2: We agree that the technical details regarding the clustering process were insufficient in the original manuscript. To address this, we have added a more detailed explanation on page 11-12, line 352-365. Specifically, we set the number of clusters to 150 for the Lego dataset, and to 10% of the total number of Gaussians for the Tanks and Temples (TnT) dataset. Furthermore, we present a comparison of stylization results under different cluster settings in Figure 6, which helps illustrate the impact of cluster quantity on the final rendering quality.

 

Comments 3: Clarify the differences from recent works, especially Chao et al. [76], perhaps through additional visualization.

Response 3: Thank you for your suggestion. While both our method and Chao et al. [76] utilize texture-based enhancements for 3D Gaussian primitives, the goals and technical focuses differ significantly. Chao et al. aims for photorealistic rendering by enriching Gaussians with learnable texture maps to better reconstruct real-world appearances. In contrast, our method is designed for non-photorealistic rendering (NPR), emphasizing artistic stylization through brushstroke-aware texture mapping and shape deformation. We also incorporate clustering and pruning mechanisms to control the abstraction level, which is beyond the scope of [76]. Since the objectives and visual goals of the two methods are fundamentally different, we believe a direct visual comparison may not be meaningful or fair. However, we have added textual clarification of these distinctions in the revised manuscript (page 6, lines 187–197).

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper propose a novel method for stylizing 3D scenes rendered using Gaussian Splatting. The method is based on an previous pre-reconstructed Gaussian Splatting of a scene and a 2D brushstroke reference, which with the style transfer is applied. Unlike previous techniques that primarily focused on color and texture, this work introduces a way to incorporate the brushstroke-inspired geometric effects present in the reference. Their approach involves pruning and clustering the Gaussian primitives for efficiency and applying texture mapping to project the brushstroke shapes onto the 3D scene, resulting in stylized renderings with richer visual appeal and enhanced geometric expressiveness. By treating each 3D Gaussian primitive as an individual brushstroke unit, the method computes the 2D projected area of each Gaussian, converts its elliptical shape to a rectangular bounding box, and applies texture mapping to project a style texture onto these areas. This allows each Gaussian to inherit the specific shape and detail of a brushstroke texture. This approach provides structurally meaningful and visually appealing brushstroke-aware stylization, capturing the distinctive local structures that define an artist's personal style, moving beyond methods that only stylize surface color and texture, and overcoming previous approaches to 3D style transfer which have generally been limited to color or texture modifications, and which often lacking an understanding of artistic shape deformation. These traditional methods primarily focus on global features, such as texture and color, while neglecting local geometric attributes like brushstroke shapes.

The presentation is well written, easy to follow and technically sound. I would only invite the authors to further a few aspects of their presentation:

1. For complex pre-constructed datasets, its Gaussian Splatting pruning approach necessitates manual adjustment of the number of clusters . This manual process is crucial because different cluster settings profoundly impact the rendered artistic style, ranging from coarser, abstract results with larger clusters to finer details with smaller ones, highlighting the need for an automatic determination method for improved usability. Please provide cues about how to address this limitation.

2. The inherent symmetry of Gaussian Splatting primitives in left-right and up-down directions appears to be an aesthetic limitation, which can lead to inverted or inconsistent details in some rendered frames that result in artifacts when the stylization is applied to video sequences. Evaluate the impact of these artifacts in the overall results, and how can this be mitigated.

3. The method does not incorporate text-based editing, which is useful for many artists in delivering more flexible and creative workflows. Is it feasible to incorporate a NLP bot to mitigate this?

 

Author Response

Comments 1: For complex pre-constructed datasets, its Gaussian Splatting pruning approach necessitates manual adjustment of the number of clusters. This manual process is crucial because different cluster settings profoundly impact the rendered artistic style, ranging from coarser, abstract results with larger clusters to finer details with smaller ones, highlighting the need for an automatic determination method for improved usability. Please provide cues about how to address this limitation.

Response 1: Yes, we agree that manual adjustment of the number of clusters plays an important role in achieving desirable stylization results. In our experiments, we set the number of clusters to 150 for the Lego dataset, and 10% of the total number of Gaussians for the Tanks and Temples (TnT) dataset. Through extensive testing, we found that using 10% yields a good trade-off between abstraction and detail for real-world scenes, and thus we adopt it as a default value for automatic determination in such cases.

At the same time, we preserve user control over this hyperparameter, allowing them to adjust the level of abstraction or detail according to their artistic preference. A more detailed explanation has been added to pages 11–12, lines 352–365 of the revised manuscript.

 

Comments 2: The inherent symmetry of Gaussian Splatting primitives in left-right and up-down directions appears to be an aesthetic limitation, which can lead to inverted or inconsistent details in some rendered frames that result in artifacts when the stylization is applied to video sequences. Evaluate the impact of these artifacts in the overall results, and how can this be mitigated.

Response 2: Thank you for your valuable suggestion. We have added Figure 9, which presents unsuccessful stylization examples. A potential solution is to add a consistency constraint during rendering. For example, the orientation of each Gaussian Splatting primitive in the previous frame can be tracked and used as a reference to maintain consistent orientation in the subsequent frame. We added discussion of this issue in the revised manuscript on page 14, lines 407–418.

 

Comments 3: The method does not incorporate text-based editing, which is useful for many artists in delivering more flexible and creative workflows. Is it feasible to incorporate a NLP bot to mitigate this?

Response 3: Thank you for your comment. Unfortunately, integrating an NLP-based editing system is not feasible under our current rendering pipeline, as it is non-differentiable. Our brush-based rendering approach replaces standard Gaussian Splatting primitives using a texture mapping process, which introduces discrete operations that break differentiability. As a result, even if we incorporate an NLP module, we would not be able to backpropagate gradients to update either the original 3D Gaussian primitives or the brushstroke stylization itself. Addressing this limitation would require a fundamental redesign of the pipeline to support end-to-end differentiability.