Flow-CLIP: Image-Conditioned Residual Scoring for CompositionalImage-Text Matching
Flow-CLIP refines image-text alignment for compositional matching by combining Global Compatibility Calibration (GCC) with Image-Conditioned Evidence Refinement (ICER). It addresses sentence-level alignment's weakness in binding objects and attributes differently. By using image context to adjust token evidence, the method improves accuracy on complex matching tasks, showing local refinement is crucial for resolving compositional ambiguities.