Practical overview
When it fits
Use it when you have suitable single-person monocular footage and need synthesized multiview sequences as input to a separate 4DGS workflow.
Who it is for
- Researchers and practitioners preparing monocular human footage for 4D Gaussian Splatting workflows.
- Developers who need configurable target views and camera coverage from a single-camera video.
- Users who want a local command-line workflow or a GUI for inference and visualization.
Documented uses
- Generate multiple viewpoint videos from a single-camera recording of a person.
- Choose the number and placement of target views, including orbit and layered-pitch layouts.
- Prepare multiview video and camera metadata for a separate 4DGS reconstruction stage.
- Inspect generated results through the provided GUI.
Requirements & limits
The repository is public and specifies the Apache-2.0 license. The documented workflow is self-hosted and uses a local Python environment.
- Input footage should be clear, portrait 9:16, at least 1080p, at least 121 frames, and show one person in a full-body or upper-body shot.
- The README reports 22 GB peak CUDA memory and 27 seconds per 121-frame video on a single RTX 4090; other hardware may differ.
- The project generates multiview video and camera data for downstream reconstruction; the README does not describe 4DGS reconstruction as part of this tool.
- The repository documents both command-line inference and a GUI for inference and visualization. Optional FlashAttention-3 or SageAttention installation is listed for faster inference.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the local Python environment
Clone 4DAnyone and install its documented environment and dependencies.
Before you start
- Git
- Conda
- Python 3.11 environment
- Access to the 4DAnyone GitHub repository
Steps
Clone the repository and change into its directory.
Initialize the GVHMR submodule with the documented Git command.
Create and activate the Conda environment using Python 3.11.
Install the dependencies from requirements.txt.
Expected outcome
The documented repository environment and dependencies are installed; inference steps are provided separately in the README.
Guide limitations
- This setup guide does not establish that installation succeeds on a particular machine.
- The README says missing models and examples download automatically on first use or can be downloaded manually.
Generate a 24-view full-orbit sequence
Run the documented inference configuration for 24 target views around a full orbit.
Before you start
- The repository environment and dependencies are installed.
- A suitable input video: one person in a full-body or upper-body shot, clear footage with no large camera movements, at least 1080p, portrait 9:16, and at least 121 frames.
- CUDA-capable GPU; the repository reports its performance on an RTX 4090.
Steps
Choose the input video and an output directory.
Run inference.py with the video_path, output_dir, and views_per_layer set to 24, as shown in the repository README.
Review the generated target-view videos and camera metadata in the output directory.
Expected outcome
A multiview output directory containing generated target videos and camera information, intended to support downstream 4DGS reconstruction.
Guide limitations
- The README presents this configuration as a dense 360-degree orbit suitable for 4DGS; it does not document the downstream reconstruction procedure.
- Reported performance is hardware-dependent and is not a compatibility guarantee.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- ant-research/4DAnyone · Official homepage title · retained in its source language