Creative Intelligence Scorer

Architecture

Input image (224×224 RGB)
        ↓
[FROZEN] CLIP-ViT-B/32 (openai/clip-vit-base-patch32)
        ↓  512-dim embedding
Projection: Linear(512→256) → ReLU → Dropout(0.2)
        ↓  256-dim shared representation
  ┌─────────────────────┐
  ↓                     ↓
CTR head             Fatigue head
Linear(256→1)        Linear(256→2)
Sigmoid              Weibull params (log_scale, log_shape)

Training data

• Meta Ad Library (Apify scrape): 3,502 real ad images — gaming, ecommerce, finance verticals
• PIL-generated synthetic ads: 18,746 images with rule-based CTR and half-life labels
• Total: 22,248 images | 80/10/10 train/val/test split

Metrics (test set)

Limitations

• CTR labels are proxy scores, not real click-through rates — derived from ad activity signals, not A/B test data.
• GradCAM is a spatial approximation — CLIP's pooler_output discards spatial structure; the 16×16 heatmap is gradient-weighted feature attribution on the projection layer, not true spatial GradCAM.
• Trained on a dataset with known label imbalance (wear-out >> cut-out).

Intended use