Regarding the poor performance of 3DGS that I observed when using your pre-trained model.

[CypressLi01]

Thanks for your work first. But when I tried to reproduce it, I found the 3DGS modeling performance was really poor. I used the two living room images you provided. There were errors during reproduction, and I only made two changes: set self.psnr = PeakSignalNoiseRatio(data_range=1.0) in PipelineMultiView(LightningModule) and removed transpose(1, 2) in cuRoPE2D(torch.nn.Module). The results are shown in the figure, with a huge gap from the baseline. Could you please offer some solutions?

[CypressLi01]

I conducted tests on the validation dataset using the weights you provided, but the quality of the rendered images generated by the prediction is extremely poor. I would like to figure out the possible causes of this issue and hope you could provide some guidance or troubleshooting tips.