AI Quality & Culling Engine
Multi-Axis AI Quality Scoring Framework
Deep breakdown of the 8 weighted scoring dimensions and GPT-5-mini evaluation contracts.
Technical Evaluation Dimensions
BestShots evaluates photographs using a multimodal evaluation model tuned specifically for professional photographer criteria. Rather than relying on a single generic score, each frame is analyzed across 8 distinct axes:
| Evaluation Axis | Default Weight | Key Heuristics |
|---|---|---|
| Primary Subject Focus | 0.25 | Eyelash and pupil sharpness, edge contrast, depth-of-field accuracy. |
| Subject Expression | 0.15 | Eye openness, authentic smiles, absence of awkward transitional grimaces. |
| Exposure & Dynamic Range | 0.15 | Preservation of highlight detail in wedding dresses, shadow recovery. |
| Motion Blur | 0.10 | Distinguishing intentional panning motion from accidental camera shake. |
| White Balance & Skin Tones | 0.10 | Natural skin tone rendition under mixed ambient and flash lighting. |
| Noise & Artifacts | 0.10 | High-ISO grain management and sensor noise profile. |
| Composition & Framing | 0.10 | Rule of thirds, leading lines, headroom, background distractions. |
| Cleanliness & Artifacts | 0.05 | Lens flare, sensor dust spots, chromatic aberration. |
Azure AI Foundry Integration
Inferences are executed through private Microsoft Azure AI Foundry enterprise endpoints using GPT-5-mini with reasoning-first validation schemas:
- Strict Schema Enforcement: Responses are validated against Zod schemas in
@bestshots/sharedbefore reaching database records. - Reasoning Tokens: Complex compositions leverage reasoning tokens to evaluate emotional context (such as tears of joy vs grimacing) before assigning a numerical score.
- Token Accounting: Every analysis logs exact prompt, completion, and reasoning token usage to
photo_analysesfor granular transparency.