multimodalfactbullishFrequency-aware semantic compensation in early generation stages can strengthen effective signal while maintaining structural coherence in inversion-free image editingComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02421v1