PRISM: Feed-Forward Single-Image 3D Reconstruction via Geometric Warp-Residual Modeling
PRISM proposes a feed-forward method for single-image 3D reconstruction by decomposing it into a parameter-free geometric prior and a learned residual correction. This replaces slow diffusion sampling with fast inference, allowing the model to reconstruct scenes quickly while maintaining high quality. It uses geometry to handle the bulk of the view prediction and a small network to fix remaining details.