2D-LFM: Lifting Foundation Model without 3D Supervision

Figure 1. 2D landmark lifting without 3D supervision matches supervised methods. Given 2D keypoints from single images (top), we compare against VGGT [19] which back-projects through predicted depth (middle) versus our 2D-LFM trained with 2D supervision only (bottom). VGGT produces distorted geometry despite accurate scene depth (> 100mm MPJPE), while 2D-LFM recovers correct object structure (8.1mm). Ground truth: red, Predictions: blue.