Zhensheng Yuan

Hi, You can add this API by yourself. Here is the example: [src/python_api.cu#L559](https://github.com/yzslab/instant-ngp/blob/593aa4da315514d554a37ef42c10fd827b2ca492/src/python_api.cu#L559) [scripts/run.py#L180](https://github.com/yzslab/instant-ngp/blob/593aa4da315514d554a37ef42c10fd827b2ca492/scripts/run.py#L180)

You need to tune window size (at least one BDP) to get maximum throughput. BDP = RTT * bandwidth, kcp window size = BDP / MTU ``` --sndwnd value set...

难怪最近多了几个STAR，原来是有人发到这来了

@keeno1982 两边的窗口设置可以不一样，服务端按服务器的1G带宽设置窗口，客户端按客户端100M下行设置接收窗口，20M上行设置发送窗口。 服务端窗口比客户端大没事（当然不超过服务器的上行带宽），因为服务端会根据客户端告知的接收窗口发包。 算出来的值只是理论值，一般来说可能需要往大设，因为发包多了，路上排队的数据包会增加，KCP要处理的数据包也越多，从而导致延迟增加。不过如果引起拥塞了，导致丢包率上升，那就应该往小设。 具体怎样设置，需要你自己调整，反复测量，算出来的值只是参考用。

分两种情况吧。 第一种：低pps 这种情况，按你10:3的比例，要发够10个才能触发生成冗余包的机制，也许等凑够10个，早就已经触发重传机制的，因此就轮不到FEC。 第二种：高pps 发包快，如果造成拥塞，丢包率会增大，你原来的比例肯定就不适用了

> Can you try training with the nerfacto model to see if you get similar artifacts. Hi, the nerfacto model works well, no artifacts. 

> I also observed that old version nerfstudio instant-ngp works well, but the current version does not work on the same capture. You can try 0.1.6. Thanks for your advice!...

> Could you please try running ingp with nerfstudio-data loader? Hi, I tried `ns-train instant-ngp --data dataset/Family nerfstudio-data` just now, it produced the same result that full of artifacts.

Hi, I also implemented a viewer for 4DGaussians, and it is web based: https://github.com/yzslab/gaussian-splatting-lightning#load-model-trained-by-other-implementations Run `python viewer.py 4DGaussians/outputs/lego --vanilla_gs4d`, and access `http://SERVER_IP:8080`, then you can view your scene interactively (NOTE:...

> > Hi, I also implemented a viewer for 4DGaussians, and it is web based: https://github.com/yzslab/gaussian-splatting-lightning#load-model-trained-by-other-implementations > > Run `python viewer.py 4DGaussians/outputs/lego --vanilla_gs4d`, and access `http://SERVER_IP:8080`, then you can view...