Estimated Preference Score

Capability answers one question — did the image follow the prompt? — but says nothing about whether the image looks good. Estimated Preference (EPS) is the second axis: an aesthetic human-preference score derived from a reward model, independent of prompt adherence.

Why a second axis

Following the prompt and looking good are different skills. A model can nail every instruction and still produce flat, lifeless images; another can be gorgeous but ignore half the prompt. Reporting only one number hides that trade-off.

Overall = 0.5 × Capability + 0.5 × EPS

Overall combines the two on equal footing so the leaderboard rewards models that both do what you asked and look the part. Because the axes are reported separately, you can still sort by pure capability or pure preference when your use case leans one way.