SADM-reproducibility-ARRAY-D-25-02292R1

CLIP_Similarity_Scores supports the empirical validation of the SADM (Staged AI-Driven Methodology) framework for dynamic poster design, specifically addressing the mandatory revision requested by the journal (ARRAY-D-25-02292R1). The file contains quantitative similarity metrics and human judgments for generated design assets compared to a given style reference. Key Variables: 1. Test ID: Unique identifier for each generated iteration. 2. CLIP Similarity Score (ViT-B/32): Semantic alignment between the generated asset and the style reference (range 0–1). Higher values indicate closer semantic style matching. 3. SSIM (Structural Similarity Index): Pixel‑level structural similarity (range 0–1). Expected to be relatively low for dynamic posters because structural divergence is tolerated. 4. LPIPS Perceptual Distance (AlexNet): Perceptual dissimilarity (range 0–1). Lower values are better; values below ~0.4 indicate good perceptual agreement. 5. Human Judgment: Expert acceptance decision (Accepted / Rejected), with qualitative justifications. 6. Basis for Judgment: Explanation of why an asset was accepted or rejected, often referencing the CLIP threshold (~0.8) or signs of AI hallucination. How the Data Are Used in the Manuscript: 1. To define and justify the hierarchical dual‑gate selection system: · Assets with CLIP > 0.82 are accepted as semantically aligned. · Assets with CLIP < 0.7 are rejected due to severe hallucination or semantic breakdown. · Assets between 0.7 and 0.8 are considered marginal and subject to designer override. 2. To demonstrate that the framework tolerates high perceptual distance (LPIPS > 0.6, SSIM < 0.35) as long as the semantic core remains intact. This tolerance is a deliberate design choice to encourage creative diversity and “daring visual possibilities” in dynamic posters. 3. To provide raw, transparent evidence that the iteration controller and similarity thresholds operate as described in Section 4.1.2 and the rebuttal letter. Relation to Reproducibility: These scores, along with the full generation logs and evaluation scripts, are part of the reproducibility package linked in the Data Availability Statement. Readers and reviewers can verify the reported thresholds, the effectiveness of the CLIP‑based filter, and the consistency of human judgments.CLIP相似性得分数据集（CLIP_Similarity_Scores）可为动态海报设计的SADM（分阶段AI驱动方法论，Staged AI-Driven Methodology）框架提供实证验证，专门响应了期刊（ARRAY-D-25-02292R1）提出的强制性修订要求。该文件收录了生成式设计素材与指定风格参考样本之间的定量相似性指标及人工评判结果。 关键变量： 1. 测试ID（Test ID）：每个生成迭代的唯一标识符。 2. CLIP相似性得分（CLIP Similarity Score, ViT-B/32）：生成素材与风格参考样本的语义对齐程度，取值范围为0至1，得分越高表示语义风格匹配度越佳。 3. SSIM（结构相似性指数，Structural Similarity Index）：像素级结构相似性，取值范围为0至1。由于动态海报允许存在结构差异，该指标的预期值通常较低。 4. LPIPS感知距离（LPIPS Perceptual Distance, AlexNet）：感知差异性，取值范围为0至1，数值越低效果越好；约0.4以下的数值表示感知一致性良好。 5. 人工评判（Human Judgment）：专家验收决策（通过/驳回），附带定性说明理由。 6. 评判依据（Basis for Judgment）：素材被通过或驳回的解释说明，通常会参考CLIP阈值（约0.8）或AI幻觉（AI hallucination）迹象。 该数据集在论文中的应用场景如下： 1. 定义并论证分层双门控选择系统： · CLIP得分大于0.82的素材将被认定为语义对齐，予以通过； · CLIP得分小于0.7的素材因存在严重幻觉或语义断裂，予以驳回； · CLIP得分介于0.7至0.8之间的素材被视为边缘案例，需由设计师人工复核。 2. 证明该框架允许较高的感知距离（LPIPS>0.6且SSIM<0.35），只要语义核心保持完整。这种宽容性是一项刻意的设计选择，旨在为动态海报激发创作多样性与"大胆视觉可能性"。 3. 提供原始透明的实证依据，验证迭代控制器与相似性阈值的运行逻辑，与论文4.1.2节及审稿回复函中的描述一致。 可复现性关联： 该得分数据集与完整生成日志、评估脚本共同构成了数据可用性声明中提及的可复现性套件。读者与审稿人可借此验证报告中的阈值设定、基于CLIP的过滤器有效性，以及人工评判的一致性。