Feed-forward Gaussian splatting methods aim for instant novel view synthesis. Two-stage designs separate pose estimation and reconstruction, but prior work like VGGT-X and 3R-GS found quality degradation and attributed it to noisy poses, adding learned refinement. The new paper challenges that assumption.
The authors analyze per-camera errors across 107 scenes from DL3DV and MipNeRF360, using COLMAP poses as reference. They find rotation error after alignment is small (median 1.6 degrees), but translation scale is offset by a large, dataset-consistent factor. Controlled ablations over 140 DL3DV scenes show the real obstacles are coordinate-frame inconsistency and translation-scale mismatch, both correctable deterministically.
With these fixes, a frozen VGGT driving a frozen TokenGS achieves 22.47 PSNR, within 0.51 dB of COLMAP reference, with no additional training, when poses are estimated from all scene views. In the sparse setting with only four context images under the standard pose-free protocol, the same pipeline matches that all-view quality, yielding a fully COLMAP-free pipeline without any refinement step.