Compatibility Issue: Poor 2DGS Training Results Using FastVGGT-Generated COLMAP Data #23

I’d like to sincerely thank the authors for open-sourcing this excellent and innovative work.
Recently, I’ve been attempting to integrate FastVGGT with 2D Gaussian Splatting (2DGS), hoping to leverage the strengths of both—FastVGGT’s efficient reconstruction capabilities and 2DGS’s high-quality rendering—to achieve better geometric reconstruction results. However, I’ve encountered a specific issue and would like to seek advice from the community.
When I use COLMAP-format data (including cameras.bin, images.bin, and points3D.bin) generated by FastVGGT。as input for 2DGS training, the final reconstruction quality falls significantly short of expectations. Specifically, the geometric structure lacks clarity, and the visual rendering is poor. The difference is quite notable when compared to 2DGS results using data generated via the traditional COLMAP pipeline.
To troubleshoot the root cause, I’ve already tried the following steps:
Confirmed that FastVGGT stably outputs files complying with COLMAP standards, with no format gaps or errors.
Verified that 2DGS can normally load the data from FastVGGT and start the training process smoothly, with no abnormalities in data parsing.
Used the same set of input images, processed them separately via the traditional COLMAP pipeline and FastVGGT pipeline, then fed both datasets into 2DGS for training. The comparison confirmed a clear difference in reconstruction quality between the two.
I would greatly appreciate community input: Has anyone successfully achieved smooth integration of FastVGGT and 2DGS while getting satisfactory reconstruction results?