GaussCast: Dependency-Aware Shared Block Retrieval for Multi-User Layered 3D Gaussian Splatting
GaussCast is a dependency-aware delivery framework for multi-user layered 3D Gaussian Splatting that aggregates concurrent user demands at an edge proxy to fetch shared prerequisite blocks only once, thereby significantly reducing upstream traffic and improving startup latency and rendering quality compared to independent per-user delivery strategies.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine standing in a vast, digital museum where every exhibit is a perfect, three-dimensional reconstruction of a real place, built from millions of tiny, glowing points of light. This technology, known as 3D Gaussian Splatting, allows people to walk through these scenes on their phones or headsets, looking around in any direction with stunning clarity. However, these digital worlds are enormous. Even with the best compression tricks available today, a single city block or a large building interior can be gigabytes in size—far too big to download completely before you start exploring. If you tried to load the whole thing at once, you would wait minutes or even hours, defeating the purpose of an instant, interactive experience. To solve this, engineers have developed "progressive" delivery, where the system sends only the parts of the scene you are currently looking at, loading more details as you move. But this approach hits a wall when many people explore the same digital space at the same time, such as students in a virtual classroom or tourists on a shared tour.
When a group of people explores a large digital scene together, they often look at different corners of the room, yet they all rely on the same foundational data to make the image appear at all. Think of it like a group of hikers starting from the same base camp but heading up different trails; they all need the same map of the valley floor before they can navigate the specific paths ahead. In current systems, if ten people join a session, the network often sends that same base map ten separate times, once for each person. This wastes bandwidth, clogs the connection, and slows down the start for everyone. A new system called GaussCast, developed by researchers at The Hong Kong University of Science and Technology, changes how this data is shared. Instead of treating every user as an isolated traveler, GaussCast acts as a smart coordinator at the edge of the network. It looks at what a group of users is about to see, identifies the common foundational blocks they all need, and fetches that shared base just once. It then adds only the unique, specific details each person needs on top of that shared foundation.
The researchers built this system to work with existing 3D models without needing to change how the images are created or compressed. They treat the digital scene as a collection of building blocks, where some blocks are essential prerequisites that must be in place before any higher-level details can be displayed. The core innovation is a planning process that understands these dependencies. If the system tries to send a detailed refinement block before its required base block arrives, the image remains broken and useless, wasting the data that was just sent. GaussCast prevents this by ensuring that whenever a block is scheduled for delivery, all its necessary prerequisites are included in the same plan. The system aggregates the short-term needs of all users in a session, calculates a shared "base" of blocks that will be useful to the group, and then fills the remaining data budget with individual supplements for each person. This approach respects the strict timing requirements of real-time rendering, ensuring that what arrives is not only the right data but also the right data at the right time.
To test how well this works, the team simulated a scenario where sixteen users explored three different large-scale 3D scenes, ranging from a single room to a multi-room building and an outdoor area, using real movement traces recorded from people navigating these spaces. The results showed that by fetching shared prerequisites only once, the system reduced the amount of data traveling from the main server to the local edge network by 26 percent. This efficiency translated directly into a faster experience for the users: the time it took to see the first clear image dropped by 27 percent, from 0.75 seconds down to 0.55 seconds. Furthermore, because the system avoided sending useless data that arrived too late or without its necessary parts, the visual quality of the scene improved by nearly one decibel in signal-to-noise ratio, a measurable gain in clarity. The system also proved fairer; users who were looking at less common angles did not suffer in quality just because their views were unique, as the system redistributed the saved bandwidth to ensure everyone got a good experience.
The study also explored how the system behaves under different conditions. When users were exploring completely unrelated areas with no overlap in their views, the system naturally shifted back to a standard, individual delivery mode, avoiding the overhead of trying to force a shared solution where none was needed. The researchers found that the benefits grew as more people joined the session and as their paths overlapped more, confirming that the system scales effectively with group size. They also tested the system's reliance on predicting where a user will look next. Even when the prediction was slightly off, the system's strict rules about dependencies prevented it from wasting data on details that couldn't be used, keeping the amount of wasted traffic low. The team verified that the extra computing power needed to manage these complex plans and check the integrity of the data was minimal, adding only a tiny fraction to the total data load.
This work highlights a shift in how we might deliver complex 3D content in the future. Rather than treating every user as a separate stream, the system recognizes that in shared environments, the underlying data is often the same. By coordinating the delivery of these shared foundations, GaussCast demonstrates that we can make large-scale, interactive 3D worlds accessible to groups of people without overwhelming the network or delaying the start. The researchers made their code and simulation tools available to the public, allowing others to build upon this method of dependency-aware sharing. While the current tests focused on static scenes and controlled simulations, the approach offers a promising path forward for virtual tours, collaborative design reviews, and educational environments where multiple people need to explore the same digital space simultaneously. The system does not require new hardware or a complete overhaul of existing 3D graphics pipelines; it simply adds a layer of intelligent coordination that ensures the right data arrives together, once, for everyone who needs it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.