Aholo Viewer - 3D Gaussian viewer open-sourced by Kujiale

What is Aholo Viewer

Aholo Viewer is a high-performance web rendering engine based on 3D Gaussian Splatting (3DGS), open-sourced by Kujiale. The tool uses chunk-level LOD streaming technology, allowing 3D content to be streamed smoothly across different devices like video, and can run city-level ultra-large scenes with billion-level Gaussian points on mobile devices. The product is compatible with mainstream 3DGS formats such as ply, spz, sog, splat, lcc, ksplat, and comes with LOD generation, format conversion, and voxel collision body generation tools, ready for developers to use out of the box. It is also paired with the Aholo spatial intelligence platform API, covering capabilities such as spatial reconstruction, cloud rendering, and 3D AI model generation, driving the 3D internet from "viewable" to "usable".

Main features of Aholo Viewer

• Smooth rendering of billion-level Gaussian scenes: Through chunk-level LOD and streaming scheduling, it supports second-level loading and smooth roaming of billion-level Gaussian point scenes in the browser, entering within 10 seconds on the first screen.
• Installation-free access on multiple devices: Fully adapted to phones, computers, and VR devices, no client download required, and an ultra-large 3D world can be experienced by opening the browser.
• Multi-format compatibility: Supports mainstream 3D Gaussian formats on the market such as ply, spz, sog, splat, lcc, ksplat.
• Complete toolchain: Built-in tools for high-quality LOD generation, data format conversion, 3D Gaussian collision body generation, editing, and picking.
• Physics collision system: Generates voxel collision bodies from 3DGS reconstructed assets, supports real-time collision queries such as rays, capsules, ground detection, and wall blocking, and supports walking mode and camera obstacle avoidance.
• Cloud hybrid rendering: Based on OpenUSD, 3DGS and high-fidelity Mesh coexist in the same scene and are hybrid-rendered in the same frame, streamed in real time from the cloud, so low-spec devices can also preview smoothly.
• Multi-level rendering configurations: Provides three rendering paths: quality priority, performance priority, and extreme performance, balancing image quality and performance as needed.

Technical principles of Aholo Viewer

• Chunk-based LOD Tree: Splits the original 3DGS data into multiple chunks (data blocks), generates different levels of LOD for each chunk, and performs level switching in units of chunks at runtime. Compared with Spark 2.0's splat-based LOD, the chunk granularity is coarser, cache hit rate is better, and scalability is stronger, making it convenient for splicing and updating city-level ultra-large scenes by block.
• Streaming batch fetching: Loads data according to frustum priority, controls memory and VRAM usage, and achieves efficient scheduling in large scenes.
• Rendering pipeline optimization: Reduces VRAM usage through multi-precision data structures, compresses per-frame GPU overhead through cache precomputation and on-demand passes, and improves data access efficiency by combining Morton Sort and detail culling.
• 3D Gaussian Splatting: Models the scene as billions of Gaussian ellipsoids with position, shape, color, and transparency, and achieves real-time rendering through differentiable rasterization.

How to use Aholo Viewer