BiCLIP

BiCLIP addresses the “modality gap” in Vision-Language Models like CLIP and SigLIP. By introducing a structured, bilinear transformation matrix, we achieve state-of-the-art domain adaptation with extreme parameter efficiency.

Key Highlights

• SOTA Performance: +15.2% average improvement over zero-shot CLIP across 11 benchmarks.
• Extreme Gains: Up to +42% improvement on specialized domains like EuroSAT.
• Geometric Insight: Validates that domain shift can be recovered via canonical transformations.