About progressive inpainting

[kkennethwu]

Hi,

At first, thanks for your great work! It seems that current code only supports single reference image. How can I run the progressive part? With only one reference image, the 360 scene (rg. bear) may not inpaint well.

Thx.

[JiuTongBro]

Same question.

[Johanan528]

Sry to late reply. We conducted relevant ablation experiments on the bear scene. Generally, one view is sufficient. In addition, regarding progressive inpainting, in fact, as in the example of apple completion in the paper, the entire process code is the same, except that after completing one view first, it switches to the second view and re-marks the mask that needs to be completed. Then, the completion is performed in the second view.

[RUiN-jiarun]

@Johanan528 I believe that using only one image as supervision for fine-tuning a 3DGS model can easily lead to overfitting of a specific perspective, so the current process may have potential issues. Is it necessary to use several images to simultaneously supervise fine-tuning?

[Johanan528]

In fact, we only do a very fast finetune about 100 steps, and we provide a very good starting point in geometry so we can avoid the problems that you said. For example, some previous work like luciddreamer and text2immersion also take this advantage of gaussians