How it works · Path B

Adding time

A still scene has three dimensions. A performance adds a fourth one: time. That's where things get difficult.

1 interactive demo · about 4 minutes

3D to 4D

Three ways to add time

Researchers have come up with three main ways to make splats move. Each one has its own tradeoffs.

A: Frame by frame

Build a completely separate splat scene for every frame of video. It’s simple, but at 60 frames per second it takes up an enormous amount of storage. And because each frame is trained on its own, neighboring frames don’t quite line up, so the result shimmers.

B: Bend one shape

Build one scene, then learn how to bend it over time. This is compact and works well for gentle motion. But anything that appears, disappears, or moves very fast is hard to describe as a bent version of something that was already there.

C: Time-native

Make time a property of every splat. Each one knows when it exists and how it moves during that time. New things simply get new splats, and things that leave let theirs fade out. Mirage is built on this approach, because performers do exactly the kinds of things the other two struggle with.

Try it
Three ways to add time
Approach
Keep an eye on the yellow ball. With frame by frame, it gets rebuilt from scratch every frame. With one bent shape, it has to exist the whole time, so it shows up as a smear before it arrives. With the time-native approach, its splats start to exist the moment it arrives. Mirage uses the time-native approach.
Check yourself

Two quick questions

Key ideas
  • Video adds a fourth dimension, time, to splats.
  • Frame by frame is simple but huge and flickery. Bending one shape has trouble with fast or new things.
  • Time-native splats each know when they exist and how they move. Mirage works this way.
1What's the main weakness of rebuilding a scene for every frame?
2With the time-native approach, what happens when something new enters the scene?
0 of 2 answered