Source-linked AI summary

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

Yudong Jin, Tao Xie, Qihang Zhang, Zehong Shen, Zhen Xu, Yujun Shen, Hujun Bao, Xiaowei Zhou, Yinghao Xu

arXiv:2608.20335v1cs.CV

TL;DR

4DAnyone addresses the difficulty of maintaining multiview consistency when camera-controlled diffusion must generate the many views required for 4DGS from monocular video. It combines fixed-length reference-context packing, routed target-view groups, and skeleton conditioning, and reports stronger video quality and downstream 4DGS reconstruction than prior methods. The paper also identifies scope boundaries for loose garments and responsible use.

  • Problem

    Camera-controlled video diffusion models do not maintain sufficient cross-view consistency when scaled to the tens of target views required for 4DGS reconstruction.

  • Method

    4DAnyone uses Reference Context Packing, Target Context Routing, 3D skeleton conditioning, and mixed-source training data to generate reconstruction-grade multiview videos for 4DGS.

  • Results

    4DAnyone significantly outperforms prior methods in both novel-view video quality and downstream 4DGS reconstruction.

  • Takeaways & Limitations

    The framework supports high-fidelity 4D human reconstruction from monocular videos using reconstruction-grade multiview-consistent video generation.

  • Takeaways & Limitations

    Loose garments can remain inconsistent across views because skeleton guidance is not informative for fabric moving far from the body.

Abstract

from arXiv · show

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.

1 Introduction

4DAnyone targets 4D human reconstruction from uncalibrated monocular video by generating consistent novel-view videos for 4DGS. It addresses reconstruction-scale consistency failures caused by bounded attention context with scalable reference conditioning and cross-group target routing.

  • 4D human reconstruction remains challenging because current 4DGS construction requires dense multiview video from calibrated, static camera arrays.
  • Camera-controlled diffusion models produce plausible novel views but struggle to maintain consistency across the tens of target views needed for 4DGS.
  • When target views exceed one DiT pass, splitting them into groups creates O(N) reference-context growth and disjoint target contexts that cause structural drift.
  • Reference Context Packing compresses generated reference views into a fixed-length mixed-resolution context, reducing reference-context complexity from O(N) to O(1) while preserving appearance guidance.
  • Target Context Routing rotates target-view groupings during high-noise denoising for cross-group structure sharing, then fixes adjacent groups during low-noise denoising to stabilize details.
  • 4DAnyone combines 3D-aware skeleton conditioning with the MVGameHuman dataset and light-stage and in-the-wild data to support reconstruction-grade consistency and generalization.

2 Related Work

Prior work spans camera-controlled video generation, multi-view diffusion, and 4D human reconstruction, while 4DAnyone targets reconstruction-grade novel-view consistency from monocular video. Its approach combines sparse 3D skeleton conditioning with scalable context management, and experiments report stronger consistency, reconstruction, and generalization than compared methods.

  • Multi-View Diffusion: Multi-view diffusion extends viewpoint-conditioned generation from static images to dynamic videos for generation-then-reconstruction pipelines.The cited methods include Zero-1-to-3, MVDream, CAT3D, SV3D, Zero123++, SV4D, and CAT4D.
  • 4D Human Avatar Reconstruction: Traditional dynamic-avatar reconstruction uses dense multi-view captures, whereas monocular methods reduce capture requirements but remain constrained by camera knowledge, occlusion, or appearance fidelity.DNA-Rendering exemplifies the expensive multi-camera setting with a 48-camera rig.
  • 4DAnyone: 4DAnyone generates target-view videos from monocular input using 3D skeleton cues, RCP conditioning, and TCR-based exchange among target groups before 4DGS reconstruction.Its target viewpoints are prescribed static views distributed around the subject.
  • Camera-Controlled Video Generation: Camera-controlled video methods use implicit or explicit camera conditioning, but dense geometry and camera parameters can be unreliable for in-the-wild videos.4DAnyone instead emphasizes accurate sparse geometric signals for viewpoint control.
  • Evaluation: 4DAnyone outperforms all baselines across 4DGS reconstruction, generated-video consistency, and generated-video reconstruction on DNA-Rendering and DyMVHumans.The evaluation uses held-out DNA-Rendering sequences and out-of-distribution DyMVHumans data.
  • Evaluation: Ablations associate RCP with appearance guidance, TCR with cross-group consistency, sliding routing with improved metrics, and depth buffering with better geometric consistency.In-the-wild and challenging-case evaluations also report robust target-view generation and 4DGS reconstruction.

5 Conclusion

4DAnyone reconstructs high-fidelity 4D humans from monocular videos using sparse 3D skeleton conditioning, scalable appearance guidance, and reduced cross-group drift. The authors report improved video quality and downstream 4DGS reconstruction, while noting risks from realistic human-video synthesis.

  • 4DAnyone combines sparse 3D skeleton conditioning, Reference Context Packing, and Target Context Routing for monocular 4D human reconstruction.
  • Realistic human-video synthesis poses risks of deepfake misuse, identity privacy violation, and copyright infringement.

Supplementary Material

The supplementary model details describe multiview attention, multi-scale RCP patchification, and a 3D-aware skeleton encoder that injects geometric residuals into DiT latents.

  • Multiview self-attention rearranges tokens so views at the same timestep directly attend to one another.It shares architecture and weights with the base model’s temporal self-attention, differing in token arrangement.
  • RCP 2× and 4× patchify layers use progressively larger spatial kernels and strides to reduce reference-token counts.Their kernels and strides are (1, 4, 4) and (1, 8, 8), respectively.
  • The 3D-aware skeleton encoder processes depth-buffered RGB skeleton videos through Conv3d layers and projects them to DiT-resolution residuals.The encoder uses 10 Conv3d layers followed by a 1×1×1 projection, with 32× spatial and 4× temporal downsampling.

B Dataset Details

MVGameHuman is an in-house game-engine dataset providing synchronized, high-resolution multiview human videos across many actors and virtual camera viewpoints.

  • MVGameHuman contains 38k synchronized multiview human videos rendered at 2560×1440 resolution.
  • The dataset covers 318 actors captured by 24 virtual cameras per sequence.

C Training Details

Training fine-tunes Wan2.2 across staged datasets and camera configurations, using perceptual losses, body-part-aware sampling, and progressively broader video conditions.

  • All training stages fine-tune Wan2.2-TI2V-5B at 704×1280 resolution with learning rate 1 × 10^-5 and LPIPS weight λ=0.25.The three stages take approximately 0.5, 1, and 1.5 days on 128 H20-3E GPUs.
  • Body-part-aware LPIPS sampling prioritizes full-body, face, and hand regions while retaining uniform global coverage.Sampling probabilities are 0.2 for full body, 0.2 for face, 0.1 for each hand, and 0.4 uniform.
  • Training samples camera and frame configurations while keeping target cameras multiplied by frame count approximately constant.Examples include 6 × 41 ≈ 4 × 61 ≈ 1 × 121 to maintain similar token counts per pass.
  • Stage-specific settings: Stage 1 uses foreground-only DNA-Rendering and independently sampled source ranges to decouple pose from appearance.
  • Stage-specific settings: Stage 2 adds unmasked MVGameHuman and SynCamVideo data, while Stage 3 adds Pexels and TedTalk monocular data and drops noisy finger keypoints.

D HMR Details

The HMR pipeline estimates a ground-aligned SMPL-X mesh sequence, then derives sparse 3D keypoints for skeleton conditioning and rendering.

  • GVHMR estimates a ground-aligned SMPL-X mesh sequence from the monocular video.
  • A sparse vertex-to-keypoint regressor extracts 70 Goliath-vocabulary 3D keypoints from the mesh.
  • Skeleton rendering retains body, foot, and palm-level hand keypoints while excluding face and individual finger joints.
  • Each Goliath keypoint is predicted as a convex combination of nearby SMPL-X vertices, reducing extraction to a sparse weighted sum.

E Inference Details

Inference takes multi-view skeleton conditions through target-view video generation and 4DGS reconstruction, with configurations and timings that scale camera coverage and runtime.

  • The inference pipeline generates all target-view videos from multi-view skeleton conditions before reconstructing the final 4DGS model.
  • Multi-GPU inference: Single-layer, two-layer, and three-layer configurations provide 16, 32, and 48 cameras for increasingly demanding view coverage.Sixteen cameras suit limited pitch angles, 32 cover most scenarios, and 48 accommodate complex clothing or extreme motions.
  • 4DGS reconstruction: FreeTimeGS initializes 4D Gaussian primitives through space carving and optimizes 16-camera, 121-frame sequences for 50k iterations.Optimization uses Adam with a learning rate of 1.6 × 10^-4.
  • Inference efficiency: Preprocessing takes approximately 2 minutes, multi-view generation approximately 7 minutes, and 4DGS training approximately 30 minutes.These timings use a single RTX 4090 for preprocessing and reconstruction, and a single H20 GPU for generation.

F Evaluation Details

Evaluation compares methods on matched DNA-Rendering and DyMVHumans test settings, while 4DAnyone generates reference views and routed target groups for reconstruction.

  • Evaluation setup: All methods use the same DNA-Rendering and DyMVHumans test sequences, with 16 approximately uniformly distributed cameras and 98 frames per scene.The evaluation includes 10 DNA-Rendering scenes and 3 DyMVHumans scenes.
  • Baselines: MV-Performer is evaluated zero-shot using source-view depth warped to target views with ground-truth camera parameters.
  • Baselines: TrajectoryCrafter is evaluated zero-shot using source-video depth aligned to dataset camera scale before warping.
  • Baselines: ReCamMaster is fine-tuned on the same training datasets and settings, with matching RCP and TCR for controlled conditioning comparison.
  • 4DAnyone: 4DAnyone first generates 4 uniformly spaced reference views, then produces 16 target views in four-view groups using TCR and fixed RCP context.The model generates 121 frames and truncates to the first 98 for test-sequence matching.

F.1 Ablation Details

The ablation evaluates routing schedules on eight DNA-Rendering scenes and finds that sliding early denoising steps helps until a 0.2 switching ratio, after which gains saturate.

  • Setup: The ablation uses eight DNA-Rendering scenes and omits reference views only for the w/o RCP condition.
  • Routing variants: Sliding, Random, and Strided dynamically regroup azimuth-ordered target views during denoising, while w/o TCR keeps contiguous groups.
  • Switching-time sweep: The ablation compares generated-video consistency across different numbers of sliding denoising steps.
  • Switching-time sweep: t_s/T = 0.2 is selected because increasing sliding steps improves consistency until that point, while further decreases provide no measurable gain.Further decreases can slightly degrade individual metrics, with fluctuations comparable to 4DGS optimization variation.

G Additional Results

4DAnyone can extend from a single image to a 4D avatar by first synthesizing the subject’s target motion and then generating multiview-consistent videos for 4DGS reconstruction.

  • Wan-Animate generates a monocular video of a subject performing target motion from a single image and driving motion video.
  • 4DAnyone converts the synthesized monocular video into multiview-consistent videos for reconstructing a 4DGS avatar.

H Limitations and Failure Cases

The main failure cases arise when skeleton guidance is unreliable for non-body motion or when human mesh recovery mis-estimates unusual poses, although the pipeline remains robust to many occluded or blurred inputs.

  • HMR typically predicts complete and plausible skeletons under occlusion or motion blur, so pose errors usually yield shifted but coherent humans.
  • Loose garments: Skeleton guidance fails for large flowing garments that move far from the body, producing inconsistent views and degraded reconstruction.
  • Inaccurate pose estimation: When HMR mis-estimates an unusual pose, all generated views faithfully inherit the incorrect skeleton.For an en-pointe dancer, HMR predicts flat feet instead of the correct pose.
Loading 2608.20335v1…