CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation
Researchers have introduced Curved Ray Expectation Positional Encoding (CRePE), a novel method for enhancing camera-controlled video generation. CRePE addresses limitations in existing methods by providing a Unified Camera Model-compatible encoding that accurately represents projected-path geometry, even with wide-angle and fisheye lenses. Implemented via a Geometric Attention Adapter within video Diffusion Transformers, CRePE improves camera control stability and perceptual quality, outperforming baseline methods in various metrics. AI
IMPACT Introduces a new positional encoding technique that enhances control and quality in AI-driven video generation models.