Rank all and only the provided enhanced images from best to worst. Evaluate based on:
1. fidelity and consistency with the reference image;
2. overall perceptual quality.
The reference image is for comparison only and must not appear in the ranking.
The allowed candidate labels are: {candidate_labels}.
Use each allowed label exactly once. Never add, omit, alter, or duplicate a label.
Return exactly:
Rank: <all allowed labels ordered from best to worst, separated by " > ">
Confidence: [0.00-1.00]
