NerfStudio continues to receive updates to its open-source platform. The platform now supports “Splatfacto-Big” as a third-party method. Splatfacto-Big is an enhanced version of Splatfacto designed to produce higher fidelity outputs.
Just like how nerfacto-big is a more powerful implementation of nerfacto, the same is true for splatfacto-big.
You can compare how splatfacto-big stacks up to the original implementation by Inria, here. It generally outperforms the original method, by a small margin, and seems to train faster in my experience. On a 3090 GPU, Splatfacto-Big takes about 15 minutes to train to 30,000 steps.
Running Splatfacto-Big requires approximately 12GB of VRAM, in contrast to the ~6GB required for standard Splatfacto. If you have sufficient VRAM and wish to use Splatfacto, I recommend training with Splatfacto-Big.
In addition to Splatfacto-Big, NerfStudio is introducing time as a variable in the Viser viewer. This addition allows for the direct utilization of dynamic-based projects within NerfStudio, enabling users to control the specific moment of a dynamic scene’s rendering. Users retain the ability to navigate through the scene freely.
Moreover, the Render panel now includes a default time slider to set the time for each keyframe. Before initiating a formal render, you can preview your dynamic renderings in preview mode. This is a new feature for Viser, as the legacy viewer did not support this. Each keyframe can set its own time value, overriding the default. During rendering (i.e., through the camera path JSON file), the render time will interpolate between keyframes.
These updates have gone relatively under the radar, but they are currently accessible.
Both of these contributions came from independent contributors, Splatfacto-big coming from AdamRashid96 and iSach for dynamic support. Adam is no stranger to nerfstudio, having previously made the render closest camera feature.
New to Gaussian Splatting? Start here
