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

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 .