4DAnyone: video generator — what it does and what you need to run it

4DAnyone is an open-source video generator from Ant Research that turns ordinary monocular video into a four-dimensional scene. The project is presented as work for SIGGRAPH Asia 2026 and is described by its authors as a tool for creating "anyone in 4D" from amateur footage.

What it does

As described by the authors, 4DAnyone tackles the task of generating a four-dimensional representation of a person or object from a single video shot with one camera. It is based on 4D Gaussian Splatting — a method for representing dynamic scenes that recovers motion and volume. The authors classify the project as a generative model and a video world model, meaning the tool does not simply replay the source clip but builds a scene model that can be further worked with.

Specific use cases — camera angles, camera control, changing the background or the duration of the result — are not listed by the authors in the repository description. Only the general purpose is stated: creating a 4D representation from monocular video.

What you need to run it

The code is written in Python and distributed under the Apache-2.0 license. The repository is open, and the source code is available at github.com/ant-research/4DAnyone.

According to the authors, running it requires a CUDA-capable graphics card. The repository description states that peak VRAM usage has been reduced to below 24 GB, allowing the model to run on consumer GPUs such as the RTX 4090. Elsewhere in the description, peak usage of 22 GB of CUDA memory is mentioned. These are the developer's claims, not independent measurements.

Who it suits

The project will be of interest to researchers and developers working with generative video models, 4D reconstruction, or Gaussian Splatting. The open-source code and Apache-2.0 license make it possible to study the implementation, fine-tune the model, and integrate it into your own pipelines.

For everyday use, the tool is less suitable: it requires a CUDA-capable GPU with around 22–24 GB of VRAM. Running on more modest GPUs is not supported, judging by the description. It is also worth noting that the project is at an early stage: the repository was created in August 2026, with the latest code change in early September 2026.

4DAnyone is an open-source research tool that solves a specific task: building a 4D scene from monocular video. It requires a serious CUDA-capable GPU and is aimed more at specialists than at the general user.