数据集质量消融实验设计 (Ablation Study on Dataset Quality)
这个设计非常好！将实验划分为“数据来源 (Architecture/Knowledge Base)” 和 “生成方法 (Method)” 这两个自变量，是顶级会议论文（如 CVPR, ICCV）中最经典的对比维度格式。
这种表格设计能实现“一箭双雕”的论证效果：
- 纵向对比（证明数据集牛）：对比 Standard Public Subset 和 Ours data 模块下的 MV-RAG / PartRAG。审稿人一眼就能看出，给同样的基线模型（MV-RAG, PartRAG）喂了你构建的高质量数据后，不仅渲染跑分大爆，连真正的3D几何指标 (F-Score, CD) 都在大幅提升，实锤你的数据质量。
- 横向对比（证明你的算法牛）：在 Ours data 模块内部，大家都在同一起跑线（用了你的高质量数据库）的前提下，你的方法（Ours）依然能打败 MV-RAG 和 PartRAG，这封顶了你算法层面的有效性。
为你整理好带示意数据的学术 Markdown 表格及分析话术，已经更新到此文件中：
1. 论文核心对比实验表格 (Table Design)
Table 1: Quantitative comparison on novel-view re-rendering and 3D reconstruction quality. We thoroughly decouple the impact of retrieval data quality and generation architectures. PSNR, SSIM, and LPIPS evaluate multi-view consistency via re-rendering. F-Score and Chamfer Distance (CD) evaluate explicitly reconstructed 3D point clouds.
| Retrieval Knowledge Base / Data Source | Method | PSNR (↑) | SSIM (↑) | LPIPS (↓) | F-Score (↑) | CD (↓) |
|---|---|---|---|---|---|---|
| None (Zero-shot) | MVDream | 18.52 | 0.742 | 0.285 | 0.654 | 0.042 |
| Standard public subset | MV-RAG | 21.05 | 0.815 | 0.210 | 0.721 | 0.035 |
| PartRAG | 21.48 | 0.824 | 0.201 | 0.735 | 0.032 | |
| Ours data (High-Quality) | MV-RAG | 23.10 | 0.856 | 0.165 | 0.788 | 0.024 |
| PartRAG | 23.54 | 0.862 | 0.158 | 0.795 | 0.022 | |
| Ours | 25.26 | 0.898 | 0.130 | 0.842 | 0.014 |
(注：表中的具体数值为模拟生成数据，排版时请根据你实际计算所得的跑分进行替换。F-Score 越高越好，CD(Chamfer Distance) 越小越好。)
2. 论文正文数据分析话术 (Text Analysis Reference)
在你介绍实验结果（Results / Ablation Study）的段落，可以直接采用以下叙述逻辑：
Disentangling Data Quality and Methodological Improvements. To demonstrate the critical role of our meticulously constructed dataset, we report a comprehensive dual-aspect ablation study in Table 1.
(1) The Superiority of Our Dataset: First, we observe a dramatic performance boost when switching the retrieval knowledge base from a standard public subset to our curated dataset. For instance, without changing any architectural constraints, supplying MV-RAG and PartRAG with Ours data explicitly lifts their geometric reconstruction performance (e.g., F-Score increments from 0.721 to 0.788 for MV-RAG, showcasing significantly more coherent 3D structures). This solidifies our claim that high-fidelity, viewpoint-aligned retrieval targets serve as an indispensable cornerstone for robust multi-view generation.
(2) Effectiveness of Our Proposed Method: Second, when strictly controlling the data source variable (i.e., all methods retrieving from our high-quality database), our proposed generation framework still consistently outperforms strong baselines. It establishes the state-of-the-art across both multi-view visual consistency (PSNR 25.26, LPIPS 0.130) and strict geometric alignment (CD 0.014). This verifies that beyond providing cleaner data, our designed adaptive geometric-semantic modulation (自适应几何-语义调制模块) is uniquely capable of fully exploiting multi-view references, effectively bridging 2D priors with 3D structural constraints.