To solve the limitations (please check Motivation page), we propose a GNN-based architecture that could enhance a 3DGS-based avatar in 1 feed forward. GNN-based architecture is suitable for irregular input data, e.g., point cloud, so we think it would be an interesting research direction to apply GNN-based model for our 3DGS-enhancement task.
As shown in the above figure, we will first transform a low-quality 3DGS head trained from low-resolution images to a graph. Then we feed the densified graph into our GNN-based model. Finally, we transform the output graph back into a 3DGS head, which we expect to be a high-quality 3DGS head. To train the model, we calculate multi-view renderings with high-resolution ground-truths.
For the GNN-based model architecture, we choose EdgeConv as our backbone module. We additionally incorporate residual connections to avoid over-smoothing.