Practical overview
When it fits
Use it when you need an editable NeRF training pipeline and can supply the Python, Taichi, PyTorch and compute environment it requires.
Who it is for
- Developers building or adapting NeRF training and rendering workflows.
- Computer-graphics and computer-vision practitioners reconstructing scenes from video or prepared datasets.
- Developers exploring mobile NeRF rendering or a NeRF backend for text-to-3D workflows.
Documented uses
- Train a NeRF for a real scene from video using the repository's preprocessing and training script.
- Train the included Lego or 360 v2 prepared-dataset examples.
- Inspect a trained scene through the interactive GUI documented for the Lego example.
- Adapt the rendering pipeline for mobile deployment using Taichi AOT.
- Use the implementation as a backend in the separate stable-dreamfusion text-to-3D project.
Requirements & limits
The repository is public and licensed under Apache-2.0.
- Video input requires COLMAP for preprocessing.
- The README's 360 v2 example uses a default batch size of 8192, reported to require up to 18 GB of RAM on an RTX 3090; it recommends adjusting batch size for the available hardware.
- The README reports real-time interactive rendering on specified iOS devices through its Taichi AOT deployment path, with device-specific frame rates.
- Training and mobile rendering are separate workflows: training uses the repository pipeline, while mobile deployment follows the Taichi AOT path.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Train a NeRF from a video
Preprocess a user-provided real-scene video and start NeRF training.
Before you start
- Install PyTorch, Taichi nightly and the repository requirements.
- Install COLMAP for video preprocessing.
- Place the video in the repository's data folder.
Steps
Choose the video path and place the video in the data folder.
Set the scale and video_fps parameters; the README recommends scale 16 for a real scene and says 150–200 generated images are typically sufficient.
Run ./scripts/train_from_video.sh -v {your_video_name} -s {scale} -f {video_fps}.
Expected outcome
The script preprocesses the video and starts training a NeRF for the captured scene.
Guide limitations
- Training resource needs depend on hardware; the README advises adjusting batch size to avoid GPU out-of-memory errors.
Train the prepared Lego scene
Train the repository's Lego example and open its interactive GUI.
Before you start
- Install PyTorch, Taichi nightly and the repository requirements.
- Use the documented Synthetic NeRF dataset with its folder name unchanged.
Steps
Download and unzip the Synthetic NeRF dataset, keeping the folder name unchanged.
Run ./scripts/train_nsvf_lego.sh.
Expected outcome
The script trains the Lego scene and displays an interactive GUI at the end of training.
Guide limitations
- The README's example performance measurement uses Ubuntu 20.04 and an RTX 3090 GPU.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- taichi-dev/taichi-nerfs · Official homepage title · retained in its source language