Resolving Representation Ambiguity in Feedforward Novel View Synthesis Transformer via Semantic-Spatial Decoupling
The paper addresses artifacts like grid-like distortions in novel view synthesis transformers caused by mixing semantic appearance and spatial information in a shared space. It proposes decoupling these streams into separate semantic and spatial tokens using an Independent-V attention mechanism. This separation preserves branch purity, leading to more structured representations and improved rendering quality while maintaining low inference latency.