About the training result

[daiyixiang666]

Coming from the Wonderland Repo.
It seems that the result of the render is quite different from the Wonderland and has some artifact, how many gassuian point are you predicting, and how is the performance between the Mamba Transformer and other?

[renq-mt]

The rendering quality is slightly inferior because we used AC3D (CogVideoX) as the video model without training it like Wonderland, resulting in more floaters.

actually, in Re10K, LatentLRM can achieve a PSNR of 29 (Wonderland: 27).
Only a portion of the DL3DV data was downloaded for training, and no testing was performed.
The number of Gaussian points is 0.5× the pixel compression rate, calculated as W//2 * H//2 * Views.

Using the Mamba Transformer architecture can significantly reduce GPU memory consumption by half, while also slightly improving inference speed."