Search papers, labs, and topics across Lattice.
The 4DAnyone framework reconstructs 4D human models from uncalibrated monocular videos by generating multiview-consistent videos and utilizing 4D Gaussian Splatting (4DGS). It addresses the limitations of existing video diffusion models, which struggle with maintaining consistency across multiple views due to bounded-attention-context issues. By introducing Reference Context Packing (RCP) and Target Context Routing (TCR), 4DAnyone significantly enhances video quality and 4DGS reconstruction performance, demonstrating robust generalization in real-world scenarios.
Achieving high-quality 4D human reconstruction from casual monocular videos could revolutionize applications in virtual reality and gaming.
We present 4DAnyone, a framework for reconstructing 4D humans from an uncalibrated monocular video by generating reconstruction-grade multiview-consistent videos and lifting them into 4D Gaussian Splatting (4DGS). Existing camera-controlled video diffusion models synthesize plausible novel-view videos but fail to maintain consistency when scaled to the tens of target views required for 4DGS reconstruction. We identify this failure as a bounded-attention-context problem: when target views exceed the capacity of a single DiT forward pass, they must be split into groups, exposing two coupled bottlenecks. On the reference-context side, conditioning on all previously generated views grows as $O(N)$, weakening cross-view appearance guidance. On the target-context side, disjoint groups cannot directly exchange information, causing global structural drift. 4DAnyone addresses both bottlenecks with two complementary designs: Reference Context Packing (RCP) compresses growing reference views into a fixed-length mixed-resolution context with $O(1)$ reference-context complexity, while Target Context Routing (TCR) rotates target-view groupings during denoising to share context across groups at high-noise steps and stabilize details at low-noise steps. We further build the MVGameHuman dataset using our in-house game engine and combine it with light-stage and in-the-wild video datasets for training. Experiments on DNA-Rendering and DyMVHumans show that 4DAnyone outperforms prior methods in both novel-view video quality and downstream 4DGS reconstruction, with robust in-the-wild generalization. See our project page for video results and source code: https://4danyone.github.io.
