Surg-UniWorld: Unified AI Model for Surgical Simulation

This research presents Surg-UniWorld, a groundbreaking unified surgical world model designed to advance surgical artificial intelligence and simulation. The goal is to synthesize highly realistic instrument-tissue interactions, which is crucial for training, planning, and developing autonomous surgical systems. Existing methods for controllable surgical video generation often struggle with a lack of a unified multimodal control paradigm, leading to problems like anatomical distortion, instrument appearance drift, and temporal inconsistencies when fusing diverse visual conditions. Surg-UniWorld addresses these challenges through several innovations. First, it constructs a "Hierarchical Surgical Anchor" from the initial frame's appearance and semantic masks. This anchor serves to preserve persistent scene identity, anatomical organization, and interaction boundaries throughout the simulation. Second, "Anchor-Relative Modality Experts" are introduced to interpret various visual cues—such as edge, depth, and optical-flow evidence—in relation to this shared anchor. This allows the model to capture complementary boundary, geometric, and motion information accurately. Finally, a "Multimodal Control Expert" performs a contribution-preserving, stage-wise composition of activated modality increments, generating precise control hints for the underlying video diffusion backbone. To support this, the researchers also created Cholec80-SurgWAM, a new benchmark dataset for controllable surgical video generation. Extensive experiments confirm that Surg-UniWorld consistently outperforms existing controllable video generation methods and surgical world-model baselines in terms of generation quality, temporal consistency, and multimodal controllability.