Dusting the Phantom: smart Gaussian allocation for dynamic scenes

Why Gaussians lose fast-moving objects

Dynamic 3D Gaussian Splatting has been a genuine breakthrough for creating high-fidelity volumetric video in real time. But the technology stumbles on a common scenario: when an object is both fast and small, it simply vanishes from the video. Fine textures in areas of intense motion get lost beyond recovery. And here's the paradox: neither extending training time nor adding more Gaussians actually fixes the problem.

A team from KAIST and other Korean universities (Minhyeok Roh, Minseo Lee, Jun Young Jeong, Byeonghyeon Lee, Lucas Yunkyu Lee, Jaesik Park, Eunbyung Park) put forward a radical hypothesis: the bottleneck isn't representational capacity, but the strategy used to allocate primitives. Existing methods place Gaussians based on their current positions and ignore where they actually need to move in space-time.

The solution: two complementary methods

Seed Injection — smartly planting new Gaussians

The method identifies insufficiently reconstructed dynamic regions by analyzing temporal changes in object appearance and reconstruction errors. Once such a region is found, the algorithm initializes new Gaussians at geometrically plausible 3D positions determined through multi-view observations. Crucially, the new primitives are placed independently of the current position of existing Gaussians, allowing them to cover previously lost areas of space.

Temporal Biased Relocation — time-aware repositioning

The second strategy takes temporal optimization signals and uses them to guide the spatiotemporal placement of Gaussians that move between frames. In simpler terms: the system learns where Gaussians should move to better capture the content of dynamic regions. Both methods operate within a fixed Gaussian budget — with no increase in computational cost.

What improved

Experimental evaluation on several dynamic-scene datasets showed significant improvement in reconstruction quality, especially in highly dynamic areas. The researchers achieved recovery of small, fast-moving objects and preservation of fine-grained texture details that previously turned into blurry artifacts.

The takeaway is simple: high-fidelity volumetric video doesn't always require piling on more compute. Sometimes it's enough to allocate resources more intelligently. The work was published on October 15, 2026; the methods are implemented without modifying the underlying Gaussian representation, making them applicable to existing video synthesis pipelines.

Full materials, including a project page with visualizations and the source code, are available at: https://minhyeok2001.github.io/Dusting-the-Phantom/ and https://github.com/minhyeok2001/DtP