CamLoop4D: Controllable 4D Scene Generation by Closing the Camera Loop between Generation and Reconstruction

Abstract Generating dynamic 3D scenes with coherent geometry and motion from a text prompt or a single image, under user‐specified camera trajectories, is a long‐standing goal in computer graphics content creation. Existing pipelines cascade a video generator and a 4D reconstructor through an error‐prone pose‐estimation step, where estimation errors compound with generation artifacts. We present CamLoop4D, a generation‐reconstruction co‐design framework that closes this loop: a camera‐conditioned video diffusion model and a 4D Gaussian reconstructor operate on the same Plücker trajectory tensor, removing pose re‐estimation by construction, and the interface is generator‐agnostic. For reconstruction, we introduce hybrid motion bases that combine six fixed se(3) generators encoding rigid‐body priors with learnable non‐rigid bases, reducing motion‐bank storage from 𝒪(B·F) to 𝒪(B) while improving fidelity; under the matched‐input protocol (all methods given the same real video and shared poses), the reconstruction module surpasses the strongest monocular baselines. CamLoop4D achieves the best PSNR/SSIM among all compared methods on the iPhone benchmark and the highest user‐preference rates against seven recent baselines in text‐to‐4D and image‐to‐4D evaluation, and it generates an 80‐frame 960 × 720 scene in ∼40 minutes on a single RTX 3090 with real‐time 20 fps playback, bringing controllable 4D content creation within reach of individual creators. Our implementation is available at https://github.com/lihao1477024473/CamLoop4D .

Authors

Institutions

Publication Details

Journal
Computer Graphics Forum
Published
2026-11-13
DOI
https://doi.org/10.1111/cgf.70621
Primary Topic
3D Shape Modeling and Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

CamLoop4D: Controllable 4D Scene Generation by Closing the Camera Loop between Generation and Reconstruction

J.S.P. Chen, H. Li
Computer Graphics Forum
3D Shape Modeling and Analysis
article

CamLoop4D: Controllable 4D Scene Generation by Closing the Camera Loop between Generation and Reconstruction

J.S.P. Chen, H. Li
article en

Abstract

Abstract Generating dynamic 3D scenes with coherent geometry and motion from a text prompt or a single image, under user‐specified camera trajectories, is a long‐standing goal in computer graphics content creation. Existing pipelines cascade a video generator and a 4D reconstructor through an error‐prone pose‐estimation step, where estimation errors compound with generation artifacts. We present CamLoop4D, a generation‐reconstruction co‐design framework that closes this loop: a camera‐conditioned video diffusion model and a 4D Gaussian reconstructor operate on the same Plücker trajectory tensor, removing pose re‐estimation by construction, and the interface is generator‐agnostic. For reconstruction, we introduce hybrid motion bases that combine six fixed se(3) generators encoding rigid‐body priors with learnable non‐rigid bases, reducing motion‐bank storage from 𝒪(B·F) to 𝒪(B) while improving fidelity; under the matched‐input protocol (all methods given the same real video and shared poses), the reconstruction module surpasses the strongest monocular baselines. CamLoop4D achieves the best PSNR/SSIM among all compared methods on the iPhone benchmark and the highest user‐preference rates against seven recent baselines in text‐to‐4D and image‐to‐4D evaluation, and it generates an 80‐frame 960 × 720 scene in ∼40 minutes on a single RTX 3090 with real‐time 20 fps playback, bringing controllable 4D content creation within reach of individual creators. Our implementation is available at https://github.com/lihao1477024473/CamLoop4D .

Computer Graphics Forum
Tsinghua University (CN)
3D Shape Modeling and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon