multimodalfactbearishAny single representation can be gamed in image generation, requiring matching against a balanced battery of encoders to prevent visible fakeness while achieving low scoresComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02375v1