4DAnyone: Creating Anyone in 4D from a Casual Monocular Video

Forum topic · 小凯 · 2026-08-22

Summary

4DAnyone is a new computer vision framework that reconstructs 4D humans from a single uncalibrated monocular video by generating reconstruction-grade multi-view consistent videos and lifting them into 4D Gaussian Splatting (4DGS). The authors identify a key failure mode in existing camera-controlled video diffusion models, called the bounded attention context problem: when the number of target views needed for 4DGS reconstruction exceeds the capacity of a single DiT forward pass, views must be split into groups. This creates two coupled bottlenecks—reference context conditioning grows with generated views, weakening cross-view appearance guidance, and disjoint target groups cannot exchange information, causing global structural drift. 4DAnyone solves this with two complementary designs: Reference Context Packing compresses growing reference views into a fixed-length mixed-resolution context, and Target Context Routing rotates target view groupings during the denoising process. Experiments on DNA-Rendering and DyMVHumans show 4DAnyone outperforms prior methods in both novel view video quality and downstream 4DGS reconstruction. arXiv: 2608.20335.

4DAnyone: Create Anyone in 4D from a Casual Monocular Video

Paper Overview

Research Area: Computer Vision (CV) Authors: Yudong Jin, Tao Xie, Qihang Zhang, Zehong Shen Published: 2026-08-22 arXiv: 2608.20335

Abstract

4DAnyone is a framework for reconstructing 4D humans from uncalibrated monocular video. It works by generating reconstruction-grade multi-view consistent videos and lifting them into 4D Gaussian Splatting (4DGS).

Existing camera-controlled video diffusion models fail to maintain consistency when synthesizing novel view videos once the number of target views reaches the dozens required by 4DGS reconstruction.

Problem: Bounded Attention Context

The paper identifies this as the "bounded attention context" problem: when the number of target views exceeds the capacity of a single DiT forward pass, views must be grouped, leading to two coupled bottlenecks:

Reference context side: conditioning grows with already-generated views, weakening cross-view appearance guidance.
Target context side: disjoint groups cannot directly exchange information, causing global structural drift.

Method

4DAnyone addresses these issues with two complementary designs:

1. Reference Context Packing — compresses the growing reference views into a fixed-length, mixed-resolution context. 2. Target Context Routing — rotates target view groupings during the denoising process.

Results

Experiments on DNA-Rendering and DyMVHumans demonstrate that 4DAnyone outperforms prior methods in both novel view video quality and downstream 4DGS reconstruction.

---

*Auto-collected on 2026-08-22*

Related: arXiv:2608.20335