PSHead: 3D Head Reconstruction from a Single Image with Diffusion Prior and Self-Enhancement

• In the introduction, the author emphasizes that "the normalization preprocessing steps of existing methods struggle in handling cases with varying scales." However, the results of PanoHead in Fig. 3 (also with shoulders) do not seem that bad. If this is a main motivation, I suggest conducting more comparisons to support it in the main paper rather than only showing a few cases in the appendix.
• The results in Fig.2 are not satisfactory, with apparent appearance and shape consistency. The $360^\circ$ videos in supplementary also show severe blur in novel views, especially in back views. In comparison, the results of Panohead are more realistic. It will be better to provide a more detailed analysis of these issues, including potential causes and ideas for improvement.
• The added SDS strategy and refinement components lead to severe efficiency degradation, nearly 1.5 hours as reported in implementations. I think the authors should conduct a comparison of runtime vs. quality metrics analysis of the trade-offs between quality improvements and computational cost.
• In ablations (section 5), the observation of the gaze direction is intuitive. However, it seems that there are no similar issues in Fig.3.

1. The paper claims to have achieved great 360 free view rendering. However, upon examining the visual results in the paper, it can be observed that the side view and back view contain excessive noise and are significantly blurrier than the front view. In comparison, it does not appear to be better than PanoHead.

2. Many techniques employed in this paper have been used in other papers with similar goals, but they don't address the limitations of these techniques. For example, in the refinement stage, it is unclear how the multi-view inconsistency of refined novel views is handled.

3. Mixed SDS. This paper utilizes three types of SDS loss. However, in Figure 4, it seems that T2V SDS only provides marginal enhancements compared to I2I SDS. Although improvements are shown in Table 3, it is not demonstrated whether T2V still performs well when the refinement stage is followed by only T2I + I2I.

4. The method section indicates that the geometry is primarily based on the SDS loss. While personalized diffusion models are mentioned, it remains unclear whether the geometry captures intrinsic details and performs better than generic SDS methods.

5. The paper reports better numerical results for novel views compared to the comparison methods. However, it is worth noting that most metrics for evaluating novel views are done in feature space rather than pixel space(such as psnr). This could explain why the novel views generated by this method appear blurry, but still achieve higher scores than the baselines.

6. The preservation of identity in the rendered avatars from novel views appears to be weak, as observed in Figure 3. In column 4, there is a noticeable change in identity.

1. The work presents incremental methods, mainly refining Head Reconstruction with predictable improvements and relying extensively on off-the-shelf modules such as 2D pre-trained diffusion models, face landmark detection, and ID recognition model (Arcface) for loss function. Specifically：

• Coarse stage: Employs DreamBooth for personalized T2I diffusion to produce a preliminary 3D-GS.
• Fine stage: Utilizes personalized T2I diffusion, landmark-guided ControlNet, and a pre-trained face refinement model (CodeFormer) .

The authors should discuss the design intuition rather than empirically constructing an engineering pipeline.

2. The complexity of the engineering pipeline, detailed in Figure 2 and Section 3, makes the work hard to follow and may hinder further exploration and industrial applications.

The authors should reduce the number of modules, focusing on core modules as the main claim.

3. PSHead lacks the capability to drive expressions.

Unlike previous works such as HeadGAP and Morphable Diffusion, PSHead does not support expression-driven animation, limiting its applicability to various downstream applications.

4. The paper omits crucial information about model parameters and reconstruction times compared to cutting-edge 3D generation works (e.g., in Tables 2 and Table 3).

5. The per-instance optimization process takes approximately 1.5 hours (refer to Implementations), indicating high computational demands.

I would appreciate it if the authors could address my concerns by providing corresponding quantitative or qualitative results based on the weaknesses and review feedback.

1. My main concern lies in the technical contributions of this paper. The authors combine multiple models, such as T2I, I2I, and T2V, to achieve state-of-the-art results. They should provide more insights regarding the use of these models in the paper.
2. The author should explain why the I2V model was not used and include an ablation study for the I2V model.
3. The optimization-based method takes a long time to create a human head Gaussian model, requiring approximately 1.5 hours on a single NVIDIA A100 (80GB) GPU, which makes it difficult to use in practical applications.