Splat4D: High-Fidelity 4D Content Creation

• Splat4D is a method for generating dynamic 4D content from monocular video by constructing temporally varying 3D Gaussian fields.
• It employs a diffusion-enhanced pipeline with multi-view generation, inconsistency identification, and an asymmetric U-Net predictor for refined video synthesis.
• The approach overcomes challenges like temporal flickering and blurry textures, enabling flexible text- and image-guided 4D editing.

Splat4D is a 2025 SIGGRAPH method for high-fidelity 4D content creation from monocular video, and also from image or text inputs via video synthesis. It targets the construction of a temporally varying 3D representation that can be rendered from novel viewpoints over time, and it does so with a diffusion-enhanced 4D Gaussian splatting pipeline that combines multi-view generation, inconsistency identification, video diffusion refinement, and a fine-tuned asymmetric U-Net Gaussian predictor. In the paper’s formulation, the final output is a 4D Gaussian representation modeled as a sequence of per-frame 3D Gaussian fields rather than a canonical Gaussian field plus an explicit deformation model (Yin et al., 11 Aug 2025).

1. Definition and scope

Splat4D addresses 4D content generation and reconstruction from monocular video. The target output is a dynamic 3D representation that supports novel-view rendering over time. The same framework is also extended to text-conditioned 4D generation, image-conditioned 4D generation, 4D human generation from an image plus motion, and text-guided 4D editing (Yin et al., 11 Aug 2025).

The paper frames monocular 4D generation as intrinsically ill-posed. A monocular video reveals only one camera trajectory, so side and back geometry remain ambiguous. The system must jointly infer appearance and motion, and must maintain consistency across frames and across viewpoints while preserving fine details. Splat4D is positioned against earlier 4D generation approaches based on SDS, 3D-aware diffusion, or video diffusion, which are described as suffering from temporal flickering, unstable motion, inconsistent geometry across viewpoints, blurry textures, and weak controllability (Yin et al., 11 Aug 2025).

Within the broader 4D Gaussian splatting landscape, Splat4D is best understood as a feed-forward 4D Gaussian reconstruction framework augmented by diffusion-based refinement. This distinguishes it from methods that focus on efficient explicit spatiotemporal representation, such as hybrid 3D-4D Gaussian allocation for dynamic scene reconstruction (Oh et al., 19 May 2025), or on style transfer over an existing 4D Gaussian backbone (Liang et al., 2024). A plausible implication is that Splat4D is less concerned with redesigning the underlying 4D Gaussian primitive than with improving spatial-temporal coherence and content quality in monocular generation.

2. System pipeline and architectural organization

The system begins from either a monocular input video or a synthesized video derived from an image or a text prompt. For video-to-4D experiments it performs foreground preprocessing with