ant-research/4DAnyone

What it solves

4DAnyone addresses the challenge of creating consistent multiview videos of a person from a single, casual monocular video. This process is essential for downstream 4D Gaussian Splatting (4DGS) reconstruction, which allows for the creation of dynamic 3D scenes from 2D video.

How it works

The system takes a single-person video as input and generates multiple synchronized videos from different camera angles (target views). It does not require input camera parameters or a static camera. Users can configure flexible camera placements, including full 360-degree orbits or specific frontal arcs, across multiple pitch layers to ensure broad coverage of the subject.

Who it’s for

This tool is designed for researchers and developers working in 4D reconstruction, computer vision, and digital human creation who need high-quality multiview data from single-view sources.

Highlights

• Flexible View Generation: Supports dozens of output views with customizable camera placement and pitch layers.
• Two Model Options: Offers a standard Base model and a distilled Turbo model for significantly faster denoising speed.
• Consumer GPU Support: Optimized to run on GPUs with less than 24 GB of VRAM (e.g., RTX 4090).
• Interactive GUI: Includes a Gradio-based interface for interactive inference and visualization using Rerun.