FocalBridge: Bridging Focal Physics and Geometric Drift for Continuous Text-to-Image Generation

While camera-aware text-to-image generation has attracted increasing attention, geometric drift remains a critical challenge in continuous focal-length generation, as small geometric deviations can accumulate across focal transitions and lead to composition distortion and scene inconsistency. Existing approaches typically encode focal length as an implicit camera condition, without explicitly regulating the geometric relationship between focal length, field of view, and scene structure during generation. To address this issue, we propose FocalBridge, a physics-guided framework that progressively recovers the focal trajectory, corrects focal-dependent geometry, and preserves high-frequency scene appearance throughout the correction process. Specifically, Recovery-LF stabilizes free-running generation by adapting the focal-control pathway to self-generated model states, while a Geometry-Control Decoder Adapter introduces explicit physical guidance through bounded geometry correction in the frozen VAE decoder. A High-Frequency Appearance Preservation objective further reduces local appearance degradation during correction. Extensive experiments show that FocalBridge substantially improves focal geometry and cross-focal scene consistency over representative baselines, with strong generalization across different scene distributions.