Research and local experiments for creating and rendering dynamic Gaussian scenes, with Ubuntu 24.04 LTS and an NVIDIA RTX 4090 as the target workstation.
The main question is: given synchronized views of an animated world, how can we reconstruct a 4D Gaussian representation, save it, and render it from different viewpoints over time?
flowchart LR
A[Multi-view images + cameras + timestamps] --> B[Prepare method-specific input]
B --> C[Train a scene or run a pretrained reconstruction model]
C --> D[Dynamic Gaussians + supporting model files]
D --> E[Images and videos]
D --> F[Compatible local interactive viewer]
Animated-world generation is supplied independently. This repository focuses on reconstruction, data preparation, representation storage, and rendering. Cloud-only proprietary services are outside scope. Downloading public code, weights, or datasets is compatible with running the computation locally.
- Follow the pretrained experiments guide: view existing scenes, compare STG and Mango-GS rendering, then try NoPo4D inference before choosing what to train.
- Read the research guide for methods, papers, downloadable assets, and license distinctions.
- Use the rendering guide for representation compatibility and additional rendering procedures.
- When ready to train, check the input-data guide and follow the local creation guide.
- For the matched Native STG contender protocol, use the contender experiments guide.
Start with splaTV's bundled scene and its local time-control patch. Next inspect the published sear_steak checkpoints, then run the bundled NoPo4D example. Record your visual assessment using the experiment template. The HUST D-NeRF bouncingballs walkthrough remains a later synthetic training option. These are experiment choices, not local quality rankings.
This repository contains research, experiment guides, and a reproducible viewer patch. It is also the workspace for future training implementations, installed environments, downloaded weights, generated 4DGS assets, and measured GPU results. Run commands here; upstream checkouts and large runtime files live beneath the Git-ignored .local/ directory.
Track our implementation, scripts, patches, environment specifications, and small experiment reports. Keep datasets, downloaded checkpoints, environments, caches, rendered images/videos, and raw logs under .local/. See the workspace setup for the layout and commands. See the execution record for native GPU preview results and remaining compatibility gates; no training or comparative benchmark is claimed.
Research was checked on 2026-09-05. The guides distinguish source-inspected behavior, author-reported results, and procedures that still need execution on the target workstation. NVIDIA device access must be checked from an authorized host terminal: a sandbox failure is not evidence of a host driver fault. The installed CUDA compiler is /usr/local/cuda-13.0/bin/nvcc; the target stack is standard Python 3.14 with uv, PyTorch 2.13.0+cu130 and torchvision 0.28.0+cu130. Follow the shared environment guide and tracked candidate specifications. No Conda or automatic downgrade is used; incompatible research implementations remain adaptation pending.
“4DGS” describes several representations, not a universal interchange format. Save the full model required by the selected renderer; a PLY file alone may omit motion networks or appearance decoders. See the compatibility table.
Open-source and research-restricted implementations are both covered, with code, dependency, and weight terms recorded separately in the license notes.