Image Generation — Methodology
Purpose
We create a fair, repeatable way to compare image-generative AI models side by side.
Test Setup
• Output size: 1024 × 1024 (1:1 square).
• Why square: neutral for both portrait and landscape use cases.
• Prompts: a rotating pool of everyday scenes (people, products, environments, text).
• Seeds: for each scene/model, we generate 5 random seeds and select the single best result.
• Settings: model’s best/maximum quality settings that run reliably on our setup.
• Post-processing: none (no upscaling, inpainting, face fixers, or manual edits). Only center-crop to 1024 × 1024 if required.
Best-of-Five Selection
We review the five random seed outputs for each scene/model and pick the best one. We record and display the winning seed (and keep the other four seeds on file for audit/repro). Selection balances prompt adherence, clarity, realism, and minimal artifacts. Tie-breakers favor cleaner subject integrity and lighting.
Scoring Characteristics (Images)
• Prompt adherence
• Ground appearance (contact with ground, shadows/reflections)
• Lighting & shadows
• Human/subject integrity (faces, hands, proportions; or object geometry)
• Text fidelity (when applicable)
• Style control
• Artifacts check (extra fingers, warps, banding, smudged text)