I noticed in Table 2 of your CVF version paper (https://openaccess.thecvf.com/content/CVPR2024/papers/Liang_LucidDreamer_Towards_High-Fidelity_Text-to-3D_Generation_via_Interval_Score_Matching_CVPR_2024_paper.pdf) that you conducted the GPTEval3D benchmark. However, since GPTEval3D requires rendering normal maps for evaluation, I am curious about how you render the normal maps for this evaluation.