4D generation synthesizes dynamic 3D scenes from conditions such as text or images. Existing methods either reconstruct generated RGB videos with a separate 4D model or adapt a particular video generator to predict geometry directly.
Latent video generation relies on autoencoders to define a compact space in which generative models operate. Although video autoencoder architectures have evolved substantially, their latent spaces are still optimized primarily for pixel-level reconstruction and provide limited high-level semantic organization.
Dream4D is a new framework for generating spatiotemporally coherent 4D content. It uses a two‑stage pipeline: first, few‑shot learning predicts optimal camera trajectories from a single image; second, a pose‑conditioned diffusion process creates geometrically consistent multi‑view sequences that are converted into a persistent 4D representation. The method uniquely combines rich temporal priors from video diffusion models with geometric awareness from reconstruction models, achieving higher quality metrics such as mPSNR and mSSIM compared to existing approaches.
arXiv:2609.35734v2 Announce Type: replace Abstract: Novel view synthesis from sparse images must reconcile faithful reconstruction of observed regions with plausible completion of unseen content, whi...
arXiv:2608. 19556v1 Announce Type: cross Abstract: Streaming autoregressive diffusion models enable real-time, long-horizon video generation, but their training objectives optimize local frame prediction rather than the geometry and dynamics of a coherent world: long rollouts accumulate geometric drift and degrade into static or unnatural motion.
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.
arXiv:2608. 20335v1 Announce Type: new Abstract: 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).
VideoTok4D introduces a 4D‑aware video tokenizer that transforms videos into compact tokens representing a dynamic 3D world. It employs spatiotemporal disentanglement to separate static and dynamic content, a track‑aware dynamic attention mechanism for motion consistency across views, and a diffusion prior (Co4DGen) for efficient 4D scene generation. Experiments show state‑of‑the‑art results with up to four orders of magnitude less storage than dense 4D representations, and shorter diffusion sequences for faster generation.
The paper introduces latent spatial memory, a 3D cache that stores scene information directly in diffusion latent space, eliminating the need for pixel-space reconstruction. It presents Mirage, a framework that lifts latent tokens into 3D using depth-guided back‑projection and queries the memory via latent‑space warping, achieving significant speed and memory gains. Experiments demonstrate up to 10.57× faster video generation, a 55× reduction in memory usage, and state‑of‑the‑art performance on WorldScore and strong reconstruction on RealEstate10K.
The paper introduces 4DGS-Fixer, an iterative refinement framework that uses a video diffusion model to enhance sparse-view 4D Gaussian Splatting for dynamic scene synthesis. It first fuses multi-view depth maps into dense point clouds for better geometric initialization, then applies a pretrained video restoration model to refine rendered sequences, providing pseudo-supervision for further refinement. Experiments on a benchmark dataset show the method outperforms existing baselines, achieving nearly a 2 dB PSNR improvement.
arXiv:2512. 05672v2 Announce Type: replace-cross Abstract: Recent approaches in controllable novel view video generation often rely on fine-tuning pre-trained Video Diffusion Models (VDMs).
Streaming autoregressive diffusion models enable real-time, long-horizon video generation, but their training objectives optimize local frame prediction rather than the geometry and dynamics of a coherent world: long rollouts accumulate geometric drift and degrade into static or unnatural motion. Recent bidirectional approaches address this problem using rewards signals built upon 3D Gaussian-Splatting reconstruction.
arXiv:2603. 03485v3 Announce Type: replace-cross Abstract: Recent video diffusion models have achieved impressive capabilities as large-scale generative world models.