English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

4D Gaussian Splatting Explained: How Millions of Fuzzy Spheres Trick Your Eyes

Forum topic · 小凯 · 2026-03-31

Summary

4D Gaussian Splatting (4D-GS) is a CVPR 2024 technique from Huazhong University of Science and Technology and Huawei that reconstructs dynamic 3D scenes as collections of millions of anisotropic 3D Gaussians whose positions, rotations, scales, colors, and opacities are optimized over time. Unlike NeRF, which uses implicit neural networks and requires hours of training with multi-second per-frame rendering, 4D-GS adopts an explicit primitive representation combined with a learned deformation field and HexPlane 4D spatiotemporal factorization. This enables training in roughly 8-30 minutes, rendering at 30-82 FPS on a single consumer GPU, and free-viewpoint video at 800x800 to 1352x1014 resolution. The article covers the mathematics of Gaussian splatting, rasterization-based rendering with alpha compositing, comparisons to NeRF, applications in volumetric capture, VFX, VR/AR, and robotics, plus current limitations and future directions including relightable 5D extensions and generative integration.

Key points

What 4D Gaussian Splatting is

  • A scene representation that uses millions of explicit 3D Gaussians instead of an implicit neural network (NeRF).
  • Each Gaussian carries about 15 parameters: position (x, y, z), color (RGB), opacity (alpha), and a 3D covariance matrix (scale + rotation quaternion).
  • Adding the time dimension via a learned deformation field, so a single canonical set of Gaussians can morph into different states at different timestamps.
  • The pipeline

    1. Initialize Gaussians from a Structure-from-Motion (SfM) sparse point cloud. 2. Encode a 4D query (x, y, z, t) using HexPlane, which decomposes the 4D spacetime volume into six 2D feature planes: XY, XZ, YZ, XT, YT, ZT. 3. Predict per-Gaussian deformation (position, rotation, scale offsets) at time t. 4. Rasterize the deformed Gaussians onto the image plane (splatting), sort by depth, and alpha-composite. 5. Optimize via differentiable rendering using L1 + SSIM loss, with adaptive densification: split Gaussians whose gradients are too large, prune low-opacity ones.

    Why it beats NeRF

    | Method | Representation | Training | Rendering | |---|---|---|---| | NeRF | Implicit MLP | 20-48 hours | <1 FPS | | 3D-GS (static) | Explicit Gaussians | 10-20 min | 100+ FPS | | 4D-GS (dynamic) | Deformed Gaussians + HexPlane | 8-30 min | 30-82 FPS |

    Speed comes from GPU-friendly rasterization instead of per-pixel ray marching. Reported PSNR is on par with or better than NeRF (25-32 dB) at resolutions up to 1352x1014, with roughly 3-4.5 GB VRAM usage.

    Applications

  • Free-viewpoint video for sports, concerts, and replays.
  • VFX: Framestore used 4D-GS in the Superman film for volumetric performance capture.
  • VR/AR and metaverse: real-time exploration of captured scenes.
  • Robotics: Splat-Nav achieves real-time 25 Hz pose estimation and path planning on Gaussian maps.
  • Digital humans, virtual idols, telepresence.
  • Limitations

  • Capture requires dozens of synchronized cameras and controlled lighting; not a phone-shot pipeline yet.
  • Topological changes (tearing, splitting) and severe topology changes are hard.
  • Semantic editing (resize a dog, recolor clothing) requires extra networks.
  • Storage of millions of Gaussians reaches hundreds of MB to several GB; compression and streaming remain open.
  • Future directions

  • 5D: separate illumination so scenes can be relit (day/night, HDR envmaps).
  • 6D: factor out material properties (roughness, metalness, IOR) for physically based rendering.
  • Coupling with diffusion models to generate 4D Gaussian scenes directly from text.
  • How to try it

  • Official code: https://github.com/hustvl/4DGaussians
  • Project page: https://guanjunwu.github.io/4dgs/
  • Paper: https://arxiv.org/abs/2310.08528
  • WebGPU viewer: https://github.com/Scthe/gaussian-splatting-webgpu
  • Hardware: RTX 3090/4090 (24 GB) recommended; RTX 2080 minimum.
  • Bottom line

    4D Gaussian Splatting shifts the field from implicit neural fields to explicit geometric primitives, achieving real-time dynamic scene rendering with modest hardware. It underpins next-generation free-viewpoint video, volumetric VFX, robotics, and is a likely building block for generative, relightable, and eventually text-to-4D worlds.

    ---

    References

  • Original Corridor Crew video: https://www.youtube.com/watch?v=X8yRlA7jqEQ
  • 4D-GS paper: https://arxiv.org/abs/2310.08528
  • 4D-GS project page: https://guanjunwu.github.io/4dgs/
  • 3D-GS (SIGGRAPH 2023): https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/
  • Splat-Nav (robotics): https://chengine.github.io/splatnav/
  • AutoDub (dubbing tool used by the author): https://github.com/king33329/audoDub

Tags

#4d-gaussian-splatting#3d-gaussian-splatting#nerf#computer-vision#real-time-rendering#free-viewpoint-video#cvpr-2024#volumetric-capture

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169460